Radiomic Analysis of MRI for Assessing Response to Neoadjuvant Chemoradiotherapy in Rectal Adenocarcinoma: A Systematic Review and Metaanalysis

Abstract

Background

This study aimed to evaluate the diagnostic performance of magnetic resonance imaging (MRI)-based radiomics for predicting pathological complete response (pCR) after neoadjuvant chemoradiotherapy in patients with locally advanced rectal adenocarcinoma.

Methods

Eligible studies developed MRI-based radiomics or deep learning models to predict pCR and reported sufficient data to reconstruct 2 × 2 contingency tables. Only validation cohorts were included in the quantitative synthesis. Study quality was assessed using Quality Assessment of Diagnostic Accuracy Studies-2 and the Radiomics Quality Score. Pooled sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), and diagnostic odds ratio were estimated using a bivariate random-effects model. Hierarchical summary receiver operating characteristic (HSROC) analysis was performed.

Results

Thirty-eight studies were included. The pooled sensitivity and specificity were 0.82 (95% CI, 0.71-0.90) and 0.86 (95% CI, 0.80-0.91), respectively. The pooled PLR and NLR were 6.0 (95% CI, 4.0-8.9) and 0.21 (95% CI, 0.12-0.35), corresponding to a diagnostic odds ratio of 29 (95% CI, 14-61). HSROC analysis showed an area under the curve of 0.846. Subgroup analyses suggested improved performance for deep learning and combined clinical–radiomic models.

Conclusion

MRI-based radiomics demonstrates good diagnostic accuracy for predicting pCR after neoadjuvant chemoradiotherapy in rectal cancer, although methodological heterogeneity and limited prospective validation remain challenges.

Keywords

rectal neoplasms magnetic resonance imaging radiomics artificial intelligence pathological complete response

Introduction

Rectal cancer remains a major global health burden and represents a substantial proportion of colorectal malignancies worldwide.¹ According to GLOBOCAN 2020 data, colorectal cancer ranks among the most frequently diagnosed cancers and is a leading cause of cancer-related mortality globally.¹ A significant proportion of patients present with locally advanced rectal cancer (LARC), for which neoadjuvant chemoradiotherapy (nCRT) followed by total mesorectal excision (TME) has been established as the standard treatment strategy.²

Despite advances in multimodal therapy, tumor response to nCRT is highly heterogeneous.³ Pathological complete response (pCR), defined as the absence of viable tumor cells in the resected specimen (ypT0N0), occurs in ∼15% to 30% of patients.⁴ Importantly, patients achieving pCR demonstrate excellent long-term oncologic outcomes, including lower local recurrence rates and improved survival.⁵ These findings have supported the development of organ-preserving strategies, such as local excision or a nonoperative “watch-and-wait” approach, in carefully selected complete responders.⁶ Avoiding radical surgery may significantly reduce morbidity, including bowel dysfunction, urinary and sexual impairment, and permanent stoma formation.⁷ Therefore, accurate preoperative identification of patients who achieve pCR is of paramount importance for individualized treatment planning.⁸

Magnetic resonance imaging (MRI) plays a central role in both primary staging and posttreatment assessment of rectal cancer due to its superior soft-tissue contrast resolution and its ability to evaluate tumor extent, circumferential resection margin involvement, extramural vascular invasion, and nodal status.⁹ MRI-based tumor regression grading (mrTRG) has been proposed to assess response after nCRT.¹⁰ However, conventional qualitative MRI assessment remains limited by interobserver variability and insufficient accuracy in distinguishing complete from incomplete responders.¹¹ Thus, reliable noninvasive imaging biomarkers capable of predicting pCR before surgery remain an unmet clinical need.¹²

Radiomics has emerged as a promising quantitative imaging approach that enables high-throughput extraction of mineable features from standard medical images.¹³ These features, including first-order intensity statistics, shape descriptors, and higher-order texture metrics, may reflect tumor heterogeneity and underlying biological characteristics that are not visually appreciable.¹⁴ By integrating radiomic features with machine learning (ML) or deep learning (DL) algorithms, predictive models can be developed to estimate treatment response and prognosis.¹⁵

In rectal cancer, MRI-based radiomics has been investigated for multiple clinical applications, including tumor staging, prediction of nodal involvement, distant metastasis, molecular status, and response to neoadjuvant therapy.¹⁶ Over the past decade, numerous studies have specifically evaluated artificial intelligence (AI)-based MRI models for predicting pCR after nCRT.¹⁷ Although many individual studies report encouraging diagnostic performance, their results are heterogeneous, and the overall pooled accuracy remains unclear.¹⁸

Furthermore, radiomics research is methodologically complex and susceptible to bias, particularly in relation to feature selection, model overfitting, lack of external validation, and limited reproducibility.¹⁹ The Radiomics Quality Score (RQS) was introduced to standardize the evaluation of methodological rigor in radiomics studies.²⁰ The Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) tool allows structured assessment of bias in diagnostic accuracy studies.²¹

Although previous systematic reviews have examined AI-assisted imaging in rectal cancer, the rapid expansion of MRI-based radiomics and DL studies in recent years necessitates an updated and focused metaanalysis specifically addressing the prediction of pathological complete response.

Methods

This systematic review and metaanalysis was conducted in accordance with the PRISMA 2020 guidelines for systematic reviews of diagnostic test accuracy studies. The study protocol was prospectively registered in the PROSPERO database (International Prospective Register of Systematic Reviews; registration number: CRD420261334791). The reporting of diagnostic accuracy measures followed principles outlined in the STARD guidelines where applicable. The methodological approach was specifically designed for diagnostic test accuracy metaanalyses and incorporated hierarchical modeling of sensitivity and specificity using established bivariate and hierarchical summary receiver operating characteristic (HSROC) frameworks.

Search Strategy

A comprehensive literature search was performed in PubMed, Web of Science Core Collection, and Scopus from database inception to February 10, 2026. The search strategy combined controlled vocabulary (MeSH and Emtree terms) with free-text keywords using Boolean operators. Search terms were adapted for each database. No initial language restrictions were applied during database searching. The full electronic search strategy for each database is provided in Supplemental Table S1. In addition, reference lists of included studies and relevant review articles were manually screened to identify potentially eligible publications not captured in the electronic search. Although Embase and the Cochrane Library were included in the search strategy, no additional eligible studies were identified from these databases after deduplication.

Eligibility Criteria

Studies were considered eligible if they included patients with histologically confirmed locally advanced rectal adenocarcinoma (clinical stage T3–T4 and/or node-positive disease) who underwent neoadjuvant chemoradiotherapy followed by surgical resection. Both standard long-course chemoradiotherapy regimens and protocol-based treatment strategies were accepted. Magnetic resonance imaging had to be used as the primary imaging modality for model development. In studies employing multimodal imaging, MRI-based results were required to be reported separately.

Eligible studies were required to develop MRI-based radiomics or artificial intelligence models, including ML or DL approaches, for the purpose of predicting pathological complete response. Pathological complete response had to be defined according to histopathological examination of the surgical specimen, typically as ypT0N0 or absence of viable tumor cells. For the purposes of this metaanalysis, MRI-based radiomics was broadly defined to include conventional handcrafted radiomics, DL-based feature extraction, and hybrid models combining both approaches. This conceptual heterogeneity was acknowledged in the interpretation.

Furthermore, studies were required to provide sufficient diagnostic performance data to enable reconstruction of 2 × 2 contingency tables. Studies were required to report a validation strategy, including internal validation or external validation. Full-text articles published in English were included.

Studies were excluded if they enrolled fewer than 10 patients, lacked a validation cohort, or did not report pathological complete response as a distinct outcome. Review articles, systematic reviews, metaanalyses, case reports, editorials, and conference abstracts were excluded. Studies evaluating neoadjuvant chemotherapy alone or radiotherapy alone were not considered eligible. Additionally, studies that reported treatment response only as “responder” versus “nonresponder” without explicit pathological complete response data were excluded. Reports with overlapping patient cohorts or duplicate populations were carefully assessed, and only the most comprehensive or recent dataset was included. Studies lacking sufficient data to reconstruct contingency tables were also excluded.

To identify overlapping cohorts, we systematically cross-checked author lists, institutional affiliations, recruitment periods, and study settings. When potential overlap was suspected, the most comprehensive or recent dataset was retained. Due to the lack of author contact in most cases, overlap identification relied on detailed comparison of reported study characteristics.

Data Extraction

Data extraction was independently performed by two reviewers using a predefined standardized data collection form. For each eligible study, information regarding study characteristics was recorded, including the first author, year of publication, country of origin, study design (prospective or retrospective), and total sample size.

Clinical and treatment-related data were also collected, including details of the neoadjuvant chemoradiotherapy protocol and surgical management. The definition of pathological complete response used in each study was documented to ensure consistency with the reference standard (ypT0N0 or absence of viable tumor cells).

MRI acquisition parameters were extracted, including field strength, imaging sequences, and timing of posttreatment imaging relative to surgery. Details of the radiomics workflow were recorded, including region-of-interest segmentation strategy (two-dimensional (2D) or three-dimensional (3D); manual, semiautomatic, or automatic), feature extraction software, number and type of radiomic features, feature selection approaches, and modeling algorithms. Validation strategies (internal split-sample, cross-validation, or external validation) and the inclusion of clinical variables in combined models were also documented.

For diagnostic accuracy analysis, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were extracted. When necessary, true-positive, false-positive, true-negative, and false-negative values were reconstructed from reported performance metrics to generate 2 × 2 contingency tables for metaanalytic pooling. When multiple models or thresholds were reported within a single study, we predefined a selection hierarchy prioritizing (1) external validation results, (2) clinically relevant models, and (3) models with prespecified thresholds. This approach minimized selective reporting bias. Discrepancies between reviewers were resolved through discussion until consensus was achieved.

Quality Assessment

The methodological quality and risk of bias of the included studies were independently evaluated using the QUADAS-2 tool and the RQS framework.

QUADAS-2 was applied to assess the risk of bias and applicability concerns in diagnostic accuracy studies across four domains: patient selection, index test, reference standard, and flow and timing. Each domain was judged as having low, high, or unclear risk of bias according to predefined signaling questions. Applicability concerns were evaluated for the first three domains. The detailed assessment criteria are provided in Supplemental Table S2.

Methodological rigor of the radiomics workflow was evaluated using the RQS. The RQS, introduced by Lambin et al,²² was specifically developed to assess the robustness, reproducibility, and clinical relevance of radiomics studies. The scoring system consists of 16 items encompassing multiple stages of the radiomics pipeline, including data selection, image acquisition and segmentation, feature extraction, model development, validation, biological correlation, and clinical utility. RQS values were calculated as absolute scores and expressed as percentages of the maximum achievable score.

Quality assessment was performed independently by two reviewers. Any discrepancies were resolved through discussion until consensus was reached. When necessary, a third senior reviewer was consulted to adjudicate disagreements.

Statistical Analysis

All statistical analyses were performed using RStudio (R Foundation for Statistical Computing, Vienna, Austria) with the mada package for diagnostic test accuracy metaanalysis. For each included study, 2 × 2 contingency tables (true positives, false positives, true negatives, and false negatives) were extracted or reconstructed when necessary to calculate study-specific sensitivity and specificity.

Pooled estimates of sensitivity, specificity, positive likelihood ratio (PLR), negative likelihood ratio (NLR), and diagnostic odds ratio (DOR) were obtained using the bivariate random-effects model of Reitsma, which jointly synthesizes sensitivity and specificity while accounting for their potential correlation and between-study variability. Summary receiver operating characteristic (SROC) curves were generated using the HSROC framework, and the AUC was calculated to summarize the overall diagnostic performance of the models.

Forest plots of sensitivity and specificity were constructed to visualize interstudy variability. Between-study heterogeneity was evaluated using the variance components of the bivariate model and I² estimates derived from the Zhou–Dendukuri and Holling approaches. I² values > 50% were considered indicative of substantial heterogeneity. In addition, visual inspection of the SROC curve, along with the corresponding confidence and prediction regions, was performed to further assess variability across studies.

To explore potential sources of heterogeneity, predefined subgroup analyses were conducted according to modeling methodology (radiomics algorithms vs DL), sample size (<100 vs ≥100), radiomics feature extraction software (PyRadiomics vs other platforms), segmentation strategy (2D vs 3D), validation approach (split/internal validation vs external validation), and model type (radiomics-only models vs combined models incorporating clinical variables).

Potential threshold effects were assessed by calculating the Spearman correlation coefficient between the logit of sensitivity and the logit of (1 − specificity). A statistically significant positive correlation suggested the presence of a threshold effect.

Publication bias in diagnostic accuracy studies was evaluated using Deeks’ funnel plot asymmetry test, with P < .10 indicating significant asymmetry. All statistical tests were two-sided, and P < .05 was considered statistically significant unless otherwise specified.

Results

Literature Search and Study Selection

The literature search identified a total of 795 records across PubMed, Web of Science Core Collection, and Scopus databases. After removal of 70 duplicate records, 725 studies remained for title and abstract screening. Following initial screening, 640 records were excluded based on irrelevance to the study objective. The full texts of 85 articles were subsequently assessed for eligibility. Of these, 47 studies were excluded for the following reasons: review articles or systematic reviews (n = 12), absence of pathological complete response as an endpoint (n = 9), nonisolated MRI radiomics analyses (n = 12), insufficient diagnostic performance data (n = 6), and overlapping patient cohorts (n = 8). Ultimately, 38 studies met the predefined eligibility criteria and were included in the qualitative and quantitative synthesis. The detailed study selection process is illustrated in Figure 1. A total of 32 non-English articles were identified during full-text screening and were excluded due to language restrictions.

Figure 1.

PRISMA 2020 flow diagram illustrating the study selection process for the systematic review of magnetic resonance imaging (MRI) radiomics studies.

Characteristics of Included Studies

The 38 included studies were published between 2018 and 2026 and reflected the rapidly expanding interest in MRI-based radiomics for treatment response prediction in rectal cancer. Most investigations were retrospective in design, while a smaller proportion were prospective cohort studies. Together, these studies evaluated patients with locally advanced rectal adenocarcinoma treated with neoadjuvant chemoradiotherapy followed by surgical resection.

Across the included studies, pathological complete response (pCR) was generally defined as ypT0N0 or absence of viable tumor cells in the resected specimen, in accordance with internationally accepted pathological staging systems. Minor variations in wording were observed, but the reference standard consistently relied on histopathological examination.

Neoadjuvant treatment protocols primarily consisted of long-course chemoradiotherapy with total radiation doses typically ranging between ∼41 and 50 Gy, administered concurrently with fluoropyrimidine-based chemotherapy, with or without oxaliplatin. Surgical management was based on total mesorectal excision principles, consistent with established oncologic standards.

MRI acquisition parameters varied across institutions. Both 1.5-T and 3.0-T scanners were used, either exclusively or in combination. The majority of studies incorporated multiparametric MRI, commonly including high-resolution T2-weighted imaging and diffusion-weighted imaging, while a subset additionally used contrast-enhanced sequences. Variability in slice thickness, acquisition protocols, and timing of posttreatment imaging relative to surgery was observed, reflecting real-world heterogeneity in clinical practice.

Segmentation strategies demonstrated methodological diversity. Manual region-of-interest delineation was the most frequently applied approach, although semiautomatic and automatic segmentation methods were reported in several studies. Both 2D and 3D segmentation techniques were utilized. Radiomic feature extraction was performed using established platforms such as PyRadiomics as well as in-house or commercial software, consistent with current radiomics standards.

The number of extracted radiomic features varied substantially among studies, often exceeding several hundred features per model. Consequently, dimensionality reduction techniques—such as Pearson correlation filtering, least absolute shrinkage and selection operator (LASSO) regression, and other regularization methods—were routinely applied to mitigate overfitting risk. Most predictive models were constructed using conventional ML algorithms, including logistic regression, support vector machines, and random forest classifiers. A subset of studies employed DL architectures, reflecting the growing integration of neural networks in oncologic imaging research.

Validation strategies included internal split-sample validation, cross-validation, and external validation cohorts. Several studies developed combined models integrating clinical parameters with radiomic features, consistent with recommendations to enhance predictive robustness and clinical applicability. A detailed overview of study characteristics and modeling approaches is presented in Tables 1 and 2.

Table 1.

Summary of General Study Characteristics.

Study	Country	Study Type	No. of Institutions	No. of Patients	MRI Field Intensity	Sequences	Slice Thickness	Image Acquisition Time	Radiotherapy Dose (cumulative)	Chemotherapy Regimen	AUC
Alvarez-Jimenez et al, 2025²³	USA	Retrospective/Prospective	3	182	1.5 T, 3.0 T	T2WI	4.0 mm	NR	45-54.0 Gy	S2: Capecitabine 825-850 mg/m² BID during RT; S3: Galunisertib 150 mg BID + 5-FU or Capecitabine during RT ± mFOLFOX6/CAPEOX	0.77
Antunes et al, 2020²⁴	USA	Retrospective	3	104	1.5 T, 3.0 T	T2WI	3.0-8.0 mm	Before nCRT	45-50.4 Gy	Capecitabine (825-850 mg/m²/day)	0.712
Azamat et al, 2022²⁵	Turkey	Retrospective	1	44	1.5 T	T2WI and DWI	3.0-4.0 mm	NR	50.4 Gy	5-FU 425 mg/m²/day + leucovorin 20 mg/m²/day (weeks 1 and 5) OR capecitabine 825 mg/m² (5 days/week, 5 weeks)	0.78
Begal et al, 2024- ²⁶	Israel	Retrospective	1	22	NR	T2WI and DWI	NR	NR	NR	Neoadjuvant chemoradiotherapy (regimens not detailed)	1.00
Bellini et al, 2022²⁷	Italy	Retrospective	1	40	3.0 T	T2WI	4 mm	NR	45-54 Gy	Oxaliplatin 50 mg/m² weekly + 5-FU 200 mg/m²/day (5 days infusion)	0.82
Boldrini et al, 2022²⁸	China	Retrospective	Two continents	220	1.5 T, 3.0 T	T2WI	NR	Before nCRT	45 Gy	Oral capecitabine 1650 mg/m² (d1-7, q7); 5-fluorouracil 225 mg/m² (d1-7, q7) or CapOx (oxaliplatin 60 mg/m² d1, q7) + oral capecitabine 1300 mg/m² (d1-7, q7)	0.75
Bulens et al, 2020²⁹	Belgium	Retrospective	2	125	3.0 T	T2WI and DWI	3.0-5.0 mm	Before nCRT	45-50 Gy	Infusion of 5-fluorouracil 225 mg/m²/day + capecitabine 825 mg/m² BID	0.86
Chen et al, 2022³⁰	China	Prospective	1	137	3.0 T	T2WI, DWI, T1WI	3.0-5.0 mm	NR	NR	CapeOx regimen: oxaliplatin 30 mg/m² (day 1) + capecitabine 850-1000 mg/m² twice daily (days 1-14), repeated every 3 weeks for 2-4 cycles	0.871
Cheng et al, 2021³¹	China	Retrospective	1	193	3.0 T	T1W, T2WI, and T2FS	3.0-4.0 mm	Before nCRT	45.0-50.4 Gy	mFOLFOX6 and CapeOX	0.912
Crimì et al, 2024³²	Italy	Retrospective	1	102	1.5 T	T2WI and DWI	3.5 mm	NR	50.4 Gy	5-FU or capecitabine-based chemotherapy (concurrent with RT)	0.871
Cui et al, 2019³³	China	Retrospective	1	186	3.0 T	T2WI, CE-T1WI, and ADC	3.0-5.0 mm	Before nCRT	50 Gy	Capecitabine (800 mg/m²/day)	0.966
Feng et al, 2022³⁴	China	Prospective	Multiple	1033	1.5 T, 3.0 T	T2WI, CE-T1WI, and DWI	2.0-6.0 mm	Before nCRT	50-45 Gy	5-Fluorouracil-based regimen ± oxaliplatin	0.812
Horvat et al, 2018³⁵	Brazil	Retrospective	1	114	1.5 T, 3.0 T	T2WI and DWI	3.0 mm	Before nCRT	NR	NR	0.93
Horvat et al, 2022³⁶	USA	Retrospective	2	164	1.5 T, 3.0 T	T2WI and DWI	3.0-5.0 mm	Before nCRT	NR	NR	0.83
Hu et al, 2024³⁷	China	Retrospective/Prospective	1	1070	1.5 T, 3.0 T	T2WI and DWI	1.0-1.5 mm	NR	45-50 Gy	Concurrent capecitabine 825 mg/m² BID (5 days/week) + RT; TME after nCRT	0.80
Huang et al, 2023³⁸	China	Retrospective	3	563	3.0 T	T2WI, CE-T1WI, and DWI	Not specified	NR	50 Gy	Tumor regression graded by Mandard TRG; pCR = TRG 1 (no viable tumor cells); Response = TRG 1-3	0.799
Jang et al, 2021³⁹	Korea	Retrospective	1	466	1.5 T, 3.0 T	T1W and T2WI	3.0 mm	Before nCRT	50.4 Gy	Concurrent fluoropyrimidine	0.76
Jiang et al, 2023⁴⁰	China	Retrospective	2	127	3.0 T	T2WI and DWI	NR	NR	50 Gy	Capecitabine or oxaliplatin + capecitabine or oxaliplatin + 5-FU during RT	0.87
Jin et al, 2021⁴¹	China	Retrospective	2	622	1.5 T, 3.0 T	T1WI, CE-T1WI, T2WI, and DWI	3.0-5.0 mm	Before and after nCRT	NR	NR	0.97
Lee et al, 2021⁴²	Korea	Retrospective	1	912	1.5 T, 3.0 T	T2WI, CE-T1WI, and DWI	3.0 mm	Before nCRT	NR	NR	0.837
Lee et al, 2024⁴³	South Korea	Retrospective	1	148	3.0 T	T2WI and DWI	3.0 mm	NR	50 Gy	Concurrent 5-FU-based chemotherapy (capecitabine or equivalent during CRT)	0.94
Lu et al, 2024⁴⁴	China	Retrospective	2	249	1.5 T, 3.0 T	T2WI	Not specified	NR	50.4 Gy	Concurrent oral capecitabine 825 mg/m² BID during RT; surgery 8-12 weeks after nCRT	0.834
Miranda et al, 2023⁴⁵	Brazil	Retrospective	1	180	1.5 T	T2WI	3.0 mm	NR	50.4 Gy	5-FU + leucovorin (IV bolus days 1-5, weeks 1 and 5, concurrent with RT)	0.836
Nardone et al, 2022⁴⁶	Italy	Retrospective	3	100	1.5 T	T2WI, ADC, and DWI	NR	Before nCRT	45 Gy	Capecitabine (825 mg/m²/day)	0.92
Pang et al, 2021⁴⁷	China	Retrospective	2	275	1.5 T	T2WI	5.0 mm	Before nCRT	45 Gy	Oral or intravenous 5-fluorouracil	0.815
Peng et al, 2023⁴⁸	China	Retrospective	1	165	3.0 T	T1W, T2WI, and DWI	1 × 1 × 1 mm³	NR	45-50.4 Gy	Long-course: Capecitabine 825 mg/m² BID during RT; Short-course: Oxaliplatin 130 mg/m² day 1 + Capecitabine 1000 mg/m² BID days 1-14	0.860
Rengo et al, 2022⁴⁹	Italy	Retrospective	1	95	1.5 T, 3.0 T	T2WI	3.0-4.0 mm	Before nCRT	45 Gy	Oxaliplatin (2-h infusion 50 mg/m²), 5-FU 200 mg/m²/day, dexamethasone (8 mg), ondansetron (8 mg)	0.833
Shaish et al, 2020⁵⁰	Italy	Retrospective	2	132	NR	T2WI	3.0-8.0 mm	Before nCRT	5 Gy	Capecitabine, 5-fluorouracil, and FOLFOX	0.80
Shin et al, 2022⁵¹	Korea	Retrospective	1	898	1.5 T, 3.0 T	T2WI, ADC, and DWI	NR	Before and after nCRT	NR	NR	0.82
Su et al, 2022⁵²	China	Retrospective	1	62	3.0 T	T2WI and DWI	3.0-4.0 mm	NR	45-50 Gy	Capecitabine 825 mg/m² BID (Monday–Friday) concurrent with RT	0.979
Wan et al, 2019⁵³	China	Retrospective	2	120	3.0 T	T1W and T2WI	3.0-5.0 mm	Before nCRT	50 Gy	Capecitabine (1650 mg/m²)	0.84
Wan et al, 2021⁵⁴	China	Retrospective	1	165	3.0 T	T2WI, T1WI, and DWI	3.0-5.0 mm	Before nCRT	45-50.4 Gy	Capecitabine (825 mg/m²/day) + oxaliplatin (130 mg/m²)	0.91
Wang et al, 2024⁵⁵	China	Retrospective	3	285	1.5 T, 3.0 T	T1W and T2WI	3.0 mm	NR	50.4 Gy	Concurrent oral capecitabine 825 mg/m² BID during RT; surgery ∼6 weeks post-nCRT	0.810
Wen et al, 2023⁵⁶	China	Retrospective	1	126	1.5 T, 3.0 T	T2WI	3.0 mm	NR	45-50 Gy	Concurrent capecitabine 1650 mg/m² daily during IMRT; ± XELOX / mFOLFOX6 / FOLFOX6 regimens	0.852
Yardimci et al, 2023⁵⁷	Turkey	Retrospective	1	76	1.5 T	T2WI	3.0 mm	NR	45 Gy	Concurrent capecitabine 825 mg/m² BID during RT; TME 6-10 weeks post-nCRT	0.753
Yi et al, 2019⁵⁸	China	Retrospective	1	134	1.5 T, 3.0 T	T1WI, CE-T1WI, and T2WI	NR	Before nCRT	46-50 Gy	Capecitabine (825 mg/m²/day)	0.908
Zhang et al, 2020⁵⁹	China	Prospective	1	383	3.0 T	T2WI, T1WI, and DKI	NR	Before and after nCRT	NR	Capecitabine (825 mg/m²/day)	0.99
Zhu et al, 2022⁶⁰	China	Retrospective	1	472	3.0 T	DWI	4.0 mm	Before nCRT	41.8-50.6 Gy	Capecitabine (825 mg/m²/day)	0.924

Abbreviations: ADC, apparent diffusion coefficient; AUC, area under the receiver operating characteristic curve; BID, twice daily; CAPEOX, capecitabine plus oxaliplatin; CapeOx, capecitabine plus oxaliplatin regimen; CE-T1WI, contrast-enhanced T1-weighted imaging; DKI, diffusion kurtosis imaging; DWI, diffusion-weighted imaging; IMRT, intensity-modulated radiotherapy; mFOLFOX6, modified fluorouracil, leucovorin, and oxaliplatin regimen; MRI, magnetic resonance imaging; nCRT, neoadjuvant chemoradiotherapy; NR, not reported; RT, radiotherapy; T1WI, T1-weighted imaging; T2FS, T2-weighted fat-suppressed imaging; T2WI, T2-weighted imaging; TME, total mesorectal excision; TRG, tumor regression grade; 5-FU, 5-fluorouracil.

Table 2.

Characteristics of Artificial Intelligence-Based Predictive Models in the Included Studies.

Study	Study VOI	Segmentation	Segmentation Software	Feature Extraction Software	Imaging Features	No. Extracted Features	ICC Evaluation	Algorithm Architecture	Validation
Alvarez-Jimenez et al, 2025²³	Primary tumor region	Manual segmentation by radiologist(s)	3D Slicer + CNN	MATLAB	Texture radiomics + High-order radiomics	773	Not reported	Radiomics-LR, LDA, and Imaging	Internal validation (CV/LOO/hold-out); External validation cohort
Antunes et al, 2020²⁴	Pre-treatment primary tumor	Manual segmentation by radiologist(s)	3D Slicer v4.8.1	MATLAB R2018a	Texture radiomics + High-order radiomics	764	Not reported	Machine Learning, Random Forest, and Imaging	Internal validation (CV/LOO/hold-out); External validation cohort
Azamat et al, 2022²⁵	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	OsiriX (Pixmeo, Switzerland)	OsiriX (first-order histogram extraction)	First-order radiomics + Morphologic features	Not predefined	Not reported	Statistical, ROC threshold, and Imaging	External validation cohort; Retrospective design
Begal et al, 2024²⁶	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	Carestream PACS (coregistration)	MATLAB R2015a	High-order radiomics	120	Not reported	Radiomics-LR, Discriminant regression, and Imaging	Internal validation (CV/LOO/hold-out)
Bellini et al, 2022²⁷	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	TexRAD (TexRAD Ltd, UK)	TexRAD histogram-based TA (with Laplacian of Gaussian filters SF0-SF2)	First-order radiomics	Not predefined	Not reported	Machine Learning, Decision Tree, and Imaging	External validation cohort; Prospective design
Boldrini et al, 2022²⁸	Pre-treatment primary tumor	Manual segmentation by radiologist(s)	MODDICOM in-house radiomics platform	MODDICOM (LoG filtering + feature extraction)	First-order radiomics	Not predefined	Not reported	Radiomics-LR, Logistic Regression, and Clinical	External validation cohort
Bulens et al, 2020²⁹	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	In-house radiomics pipeline (Python-based; Gevaert Lab)	Custom image-feature pipeline (intensity, shape, texture, wavelet, and Gabor)	First-order radiomics + Texture radiomics + High-order radiomics + Morphologic features	2131	Not reported	Radiomics-LR, Logistic Regression, and Imaging	External validation cohort
Chen et al, 2022³⁰	Pretreatment primary tumor	Manual segmentation by radiologist(s)	ITK-SNAP v3.8.0	PyRadiomics (Python v3.0)	First-order radiomics + Texture radiomics	1301	ICC > 0.70	Radiomics-LR, Logistic Regression, and Clinical + Radiomics	External validation cohort
Cheng et al, 2021³¹	Pretreatment primary tumor	Manual segmentation by radiologist(s)	ITK-SNAP v3.6	PyRadiomics (Python v3.6)	First-order radiomics + Texture radiomics + High-order radiomics	1967	ICC > 0.80	Radiomics-LR, Logistic Regression, and Clinical + Radiomics	External validation cohort
Crimì et al, 2024³²	Pretreatment primary tumor	Manual segmentation by radiologist(s)	Trace4Research™ radiomic platform (DeepTrace Technologies)	Trace4Research™ (IBSI-compliant workflow; 64-bin discretization)	Morphologic features + Texture radiomics + High-order radiomics	802	Not reported	Machine Learning, Random Forest, and Imaging	External validation cohort
Cui et al, 2019³³	Pretreatment primary tumor	Manual segmentation by radiologist(s)	A.K. software (Analysis Kit, GE Healthcare)	A.K. radiomics platform (GE Healthcare; z-score normalization)	First-order radiomics + Texture radiomics	1188	ICC > 0.75	Radiomics-LR, Logistic Regression, and Clinical + Radiomics	External validation cohort
Feng et al, 2022³⁴	Pretreatment primary tumor	Manual segmentation by radiologist(s)	MRI segmentation platform (institutional workflow); CellProfiler (nucleus); VGG-19 CNN (microenvironment)	PyRadiomics v2.1.1 (MRI); CellProfiler (770 nucleus features); VGG-19 CNN (220 microenvironment features)	Texture radiomics + Morphologic features + Deep learning features	2106	Not reported	Hybrid, SVM, and Multimodal	External validation cohort; Prospective design; Retrospective design
Horvat et al, 2018³⁵	Posttreatment tumor region	Manual segmentation by radiologist(s)	ITK-SNAP v3.4.0	In-house MATLAB (v2015b) + C++ (Insight Toolkit)	First-order radiomics + Texture radiomics + High-order radiomics	Not predefined	Not reported	Machine Learning, Random Forest, and Imaging	Internal validation (CV/LOO/hold-out); External validation cohort
Horvat et al, 2022³⁶	Posttreatment tumor region	Manual segmentation by radiologist(s)	ITK-SNAP v3.4.0	CERR radiomics platform (IBSI-compliant; 32-bin discretization; C++/R pipeline)	Texture radiomics + Morphologic features	124	CCC ≥ 0.75	Hybrid, Random Forest, and Ensemble	Internal validation (CV/LOO/hold-out); External validation cohort
Hu et al, 2024³⁷	Pretreatment primary tumor	Manual segmentation by radiologist(s)	ITK-SNAP v3.4.3	PyRadiomics (3D extraction; z-score normalization; B-spline resampling)	First-order radiomics + Texture radiomics + High-order radiomics	1237	ICC ≥ 0.75	Machine Learning, SVM, and Clinical + Radiomics	External validation cohort; Prospective design
Huang et al, 2023³⁸	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	Darwin Research Platform	PyRadiomics (via Darwin platform)	First-order radiomics + Texture radiomics	Not predefined	ICCs > 0.6	Machine Learning, SVM, and Imaging	External validation cohort
Jang et al, 2021³⁹	Posttreatment tumor region	Manual segmentation by radiologist(s)	MATLAB R2020a (Deep Learning Toolbox; ShuffleNet pretrained on ImageNet)	Deep learning automatic feature extraction (CNN + LSTM; no handcrafted features)	Deep learning features	Not predefined	Not reported	Deep Learning, CNN + LSTM, and Imaging	Internal validation (CV/LOO/hold-out); External validation cohort
Jiang et al, 2023⁴⁰	Pretreatment primary tumor	Manual segmentation by radiologist(s)	Huiying Medical Research Platform	Huiying platform (PyRadiomics + custom features; first-order, shape, texture, and wavelet)	First-order radiomics + Texture radiomics + High-order radiomics	1409	Not reported	Radiomics-LR, Logistic Regression, and Clinical + Radiomics	External validation cohort
Jin et al, 2021 ⁴¹	Pre- and post-treatment primary tumor	Manual with AI-assisted segmentation	TensorFlow + Keras implementation (3D RP-Net framework)	Deep learning automatic feature extraction (3D CNN; no handcrafted radiomic features)	Deep learning features	Not predefined	Not reported	Deep Learning, 3D CNN, and Imaging	External validation cohort
Lee et al, 2021⁴²	Pretreatment primary tumor	Not clearly specified	3D Slicer (for VOI); automatic MRI co-registration pipeline	PyRadiomics (3740 features) + 3D-ResNet (MSFI embedding, 512 features)	First-order radiomics + Texture radiomics + High-order radiomics + Deep learning features	3740	Not reported	Hybrid, CNN + RF, and Imaging	Validation strategy not clearly specified
Lee et al, 2024⁴³	Pretreatment primary tumor	Manual segmentation by radiologist(s)	MEDIP Pro v2.0.0.0 (MEDICALIP, Seoul, Korea)	MEDIP Pro (PyRadiomics-based; PCA dimension reduction; min–max scaling)	Morphologic features + Texture radiomics	116	Not reported	Machine Learning, Best selected, and Imaging	Internal validation (CV/LOO/hold-out); External validation cohort
Lu et al, 2024⁴⁴	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	uAI Research Portal (United Imaging Intelligence, Shanghai, China)	uAI Research Portal (IBSI-compliant radiomics; z-score normalization; variance threshold + LASSO)	First-order radiomics + Morphologic features + Texture radiomics	2264	ICC > 0.80	Hybrid, CNN + Logistic Regression, and Clinical + Radiomics	External validation cohort
Miranda et al, 2023⁴⁵	Pretreatment primary tumor	Manual segmentation by radiologist(s)	ITK-SNAP v3.4.0	PyRadiomics (Python; Gabor filter + gray-level discretization 8/16/32 bins)	First-order radiomics + Morphologic features + Texture radiomics + High-order radiomics	2313	ICC > 0.85	Radiomics-LR, Logistic Regression, Clinical + Radiomics	External validation cohort
Nardone et al, 2022⁴⁶	Posttreatment tumor region	Manual segmentation by radiologist(s)	LifeX Software v7.2 (INSERM, Paris, France)	LifeX Software (GLCM, histogram, shape features; delta calculation [(T2-T1)/T1])	Texture radiomics + First-order radiomics	Not predefined	ICC > 0.70	Statistical, Logistic Regression, and Imaging	External validation cohort
Pang et al, 2021⁴⁷	Posttreatment tumor region	Manual with AI-assisted segmentation	PyTorch 1.8.1 (tsraU-Net implementation); ITK-SNAP for training annotations	PyRadiomics (93 features × original + 4 wavelet subbands; per-slice extraction with statistical aggregation)	First-order radiomics + Texture radiomics + High-order radiomics	474	Not reported	Hybrid, CNN + SVM, and Imaging	External validation cohort
Peng et al, 2023⁴⁸	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	Radcloud v3.1.0 (Python 3.8.1; PyRadiomics-based)	PyRadiomics (mRMR + LASSO feature selection)	First-order radiomics + Texture radiomics + High-order radiomics	1409	Not reported	Radiomics-LR, Logistic Regression, and Clinical + Radiomics	External validation cohort
Rengo et al, 2022⁴⁹	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	Standard clinical workstation (manual measurement; no dedicated segmentation software)	WEKA v3.8.4 (data mining classification; no radiomics extraction)	Morphologic features	Not predefined	Not reported	Machine Learning, Decision Tree, and Imaging	Validation strategy not clearly specified
Shaish et al, 2020⁵⁰	Pretreatment primary tumor	Manual segmentation by radiologist(s)	3D Slicer 4.0	PyRadiomics (bin width 25; wavelet, square, sqrt, log, exp, LBP filters; contour perturbation for robustness)	First-order radiomics + Morphologic features + Texture radiomics + High-order radiomics	1595	Not reported	Radiomics-LR, Logistic Regression, and Imaging	External validation cohort
Shin et al, 2022⁵¹	Posttreatment tumor region	Not clearly specified	3D Slicer v4.10	PyRadiomics v2.1.2 (z-score normalization; fixed bin width 3 for T2, 20 for ADC; voxel resampling 1 × 1 × 1 mm; wavelet + LoG filters)	First-order radiomics + Texture radiomics + High-order radiomics	1132	ICC > 0.75	Radiomics-LR, Logistic Regression, and Imaging	Validation strategy not clearly specified
Su et al, 2022⁵²	Pretreatment primary tumor	Manual segmentation by radiologist(s)	ITK-SNAP (3D segmentation); ADW 4.5 FunctionTool (IVIM measurement)	A.K. software (Analysis Kit v2.1, GE Healthcare)	First-order radiomics + Texture radiomics + High-order radiomics	1656	ICC > 0.75	Radiomics-LR, Logistic Regression, and Clinical + Radiomics	Internal validation (CV/LOO/hold-out); External validation cohort
Wan et al, 2019⁵³	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	PACS workstation (Carestream.GCRIS)	No radiomics software; manual measurement + R statistical software (lasso logistic regression)	Morphologic features	Not predefined	Not reported	Radiomics-LR, Logistic Regression, and Imaging	External validation cohort
Wan et al, 2021⁵⁴	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	Radcloud platform v3.1.0 (http://radcloud.cn)	Radcloud platform (1049 features per sequence; delta-RF = % change post vs pre)	First-order radiomics + Morphologic features + Texture radiomics + High-order radiomics	1049	Not reported	Radiomics-LR, Logistic Regression, and Multimodal	External validation cohort
Wang et al, 2024⁵⁵	Pretreatment primary tumor	Manual segmentation by radiologist(s)	MITK v2021.10 (DKFZ, Heidelberg)	PyRadiomics v3.0.1	First-order radiomics + Morphologic features + Texture radiomics + High-order radiomics	1731	ICC ≥0.75	Machine Learning, MLP, and Imaging	External validation cohort
Wen et al, 2023⁵⁶	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	MaZda v4.6.2	MaZda radiomics platform (Z-score normalization; Pearson redundancy filtering; ANOVA/Relief feature selection)	Texture radiomics + High-order radiomics	250	Not reported	Machine Learning, Boosting/GP, and Clinical + Radiomics	External validation cohort
Yardimci et al, 2023⁵⁷	Pretreatment primary tumor	Manual segmentation by radiologist(s)	3D Slicer v4.8.2	PyRadiomics (original, LoG σ=2/4/6 mm, wavelet features; 32-bin discretization; 1 × 1 × 1 mm resampling)	Morphologic features + Texture radiomics + High-order radiomics	1046	ICC ≥ 0.90	Machine Learning, Random Forest, and Imaging	Internal validation (CV/LOO/hold-out); External validation cohort
Yi et al, 2019⁵⁸	Pretreatment primary tumor	Manual segmentation by radiologist(s)	MaZda v4.6 (Institute of Electronics, Technical University of Lodz, Poland)	MaZda v4.6 (texture extraction) + MATLAB 2017a (model building)	Texture radiomics + High-order radiomics	340	Not reported	Hybrid, RF + SVM, and Clinical + Radiomics	External validation cohort
Zhang et al, 2020⁵⁹	Pre- and post-treatment primary tumor	Manual segmentation by radiologist(s)	ITK-SNAP v3.8	Keras 2.1.5 + TensorFlow 1.4.0 (Python 3.6; custom CNN)	Deep learning features	Not predefined	Not reported	Deep Learning, CNN, and Imaging	External validation cohort; Prospective design
Zhu et al, 2022 ⁶⁰	Pretreatment primary tumor	Manual segmentation by radiologist(s)	ITK-SNAP	Custom Python pipeline (histogram concatenation; PyTorch 1.4.0 for DL)	First-order radiomics	Not predefined	Not reported	Deep Learning, 2D CNN, and Imaging	External validation cohort

Abbreviations: AI, artificial intelligence; CNN, convolutional neural network; CV, cross-validation; DKI, diffusion kurtosis imaging; DL, deep learning; DWI, diffusion-weighted imaging; ICC, intraclass correlation coefficient; IBSI, Image Biomarker Standardisation Initiative; LASSO, least absolute shrinkage and selection operator; LDA, linear discriminant analysis; LOO, leave-one-out validation; LR, logistic regression; ML, machine learning; MLP, multilayer perceptron; MRI, magnetic resonance imaging; PCA, principal component analysis; RF, random forest; ROI, region of interest; SVM, support vector machine; VOI, volume of interest.

RQS and Risk of Bias Assessment

Methodological quality of the included studies was assessed using the RQS framework. The overall RQS analysis demonstrated moderate methodological rigor, with substantial variability across studies. While most investigations adequately described imaging protocols and feature extraction procedures, critical components such as phantom studies for scanner harmonization, prospective validation, cost-effectiveness analysis, and open science practices were rarely implemented.

Items related to feature reduction, internal validation, and reporting of model performance were more consistently fulfilled. However, external validation remained limited to a subset of studies, underscoring persistent challenges in the reproducibility and generalizability of radiomics models. Biological correlation analyses and prospective study designs were infrequently reported, consistent with previously identified limitations in radiomics research.

Risk of bias and applicability concerns were evaluated using the QUADAS-2 tool. In the patient selection domain, several studies were judged to have unclear or high risk of bias due to retrospective enrollment or insufficient reporting of inclusion procedures. In the index test domain, risk of bias was frequently categorized as unclear because model thresholds were not prespecified or insufficiently detailed, a common limitation in diagnostic AI research. The reference standard domain was generally rated as low risk of bias, as histopathological evaluation after surgical resection represents the accepted gold standard for pCR determination. In the flow and timing domain, some studies demonstrated unclear risk due to incomplete reporting of the interval between MRI and surgery.

Overall, while diagnostic performance was promising, methodological heterogeneity and incomplete adherence to radiomics quality standards highlight the need for standardized imaging protocols, prospective validation, and transparent reporting in future investigations. A comprehensive summary of the RQS item analysis and QUADAS-2 assessment is provided in Figures 2 and 3. The median RQS score was 11 (range: 4-18), corresponding to 30.6% (range: 11.1%-50.0%) of the maximum achievable score.

Figure 2.

Radiomics Quality Score (RQS) item analysis. Bar chart showing the distribution of achieved scores across individual RQS items. Higher scores were most frequently observed for feature reduction and validation domains, whereas items related to prospective validation, biological correlates, phantom studies, test–retest analysis, and decision curve analysis were less commonly fulfilled.

Figure 3.

Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2) risk of bias assessment. (A) The traffic light plot shows the risk of bias judgement for each individual study across the four QUADAS-2 domains (Patient Selection, Index Test, Reference Standard, and Flow and Timing). (B) The summary plot presents the overall proportion of studies rated as low (green), unclear (yellow), or high (red) risk of bias within each domain.

Metaanalysis of Diagnostic Performance

A total of 38 studies were included in the quantitative synthesis. For studies reporting both training and validation results, only data from validation cohorts were incorporated into the metaanalysis to reduce overfitting bias and better reflect real-world model performance.

Using a bivariate random-effects model, the pooled sensitivity of MRI-based radiomics for predicting pathological complete response after neoadjuvant chemoradiotherapy was 0.82 (95% CI, 0.71-0.90), and the pooled specificity was 0.86 (95% CI, 0.80-0.91). The pooled PLR was 6.0 (95% CI, 4.0-8.9), and the pooled NLR was 0.21 (95% CI, 0.12-0.35). The corresponding pooled DOR was 29 (95% CI, 14-61), indicating strong discriminatory capacity. The HSROC analysis demonstrated an overall pooled AUC of 0.846, indicating good overall diagnostic performance (Figure 4).

Figure 4.

Forest plots of pooled sensitivity and specificity. (A) Sensitivity. Pooled sensitivity estimated using a random-effects model (logit transformation, machine learning [ML]). Squares represent individual studies (size proportional to weight), lines show 95% CI, and the diamond indicates the pooled estimate with 95% CI. Heterogeneity is expressed as I². (B) Specificity. Pooled specificity estimated using a random-effects model (logit transformation, ML). Squares represent individual studies (size proportional to weight), lines show 95% CI, and the diamond indicates the pooled estimate with 95% CI. Heterogeneity is expressed as I².

However, substantial heterogeneity was observed across studies. The I² statistic indicated significant variability in pooled sensitivity (I² = 78.76%) and specificity (I² = 90.92%). Forest plots of sensitivity and specificity are presented in Figure 4. The HSROC curve (Figure 5) demonstrates a clear separation between the 95% confidence region and the 95% prediction region, reflecting considerable between-study heterogeneity and indicating that diagnostic performance may vary across different clinical settings. These findings suggest that although MRI-based radiomics models show high overall accuracy, methodological and clinical heterogeneity among studies may influence their generalizability. The Spearman correlation coefficient between logit sensitivity and logit (1 − specificity) was 0.03 (P = .86), indicating no significant threshold effect.

Figure 5.

HSROC curve (Reitsma bivariate random-effects model). The HSROC curve summarizes diagnostic accuracy across studies. Dots represent individual studies, the ellipse indicates the 95% confidence region, and the dashed line shows the prediction region. The overall discriminative performance is expressed by the area under the curve (AUC = 0.846).

Subgroup Analysis

To explore potential sources of between-study heterogeneity, predefined subgroup analyses were conducted according to six methodological and clinical factors: modeling approach (radiomics-based ML vs DL), sample size (<100 vs ≥100 patients), feature extraction software (PyRadiomics vs other platforms), segmentation strategy (2D vs 3D), validation approach (external vs internal validation), and model composition (radiomics-only models vs combined models incorporating clinical variables).

Across subgroups, MRI-based radiomics models consistently demonstrated moderate to high diagnostic performance in predicting pathological complete response. However, variations in pooled sensitivity, specificity, and AUC were observed depending on modeling strategy and validation design.

Models developed using DL architectures showed comparable or slightly improved diagnostic performance relative to conventional radiomics-based ML approaches, although heterogeneity remained substantial. Studies with larger sample sizes (≥100 patients) tended to demonstrate more stable pooled estimates, suggesting improved model robustness and reduced small-study effects.

Regarding feature extraction software, studies employing standardized platforms such as PyRadiomics demonstrated similar diagnostic performance compared with those using other software tools, indicating that performance differences were likely attributable to modeling strategy and validation rigor rather than software selection alone.

3D segmentation approaches generally yielded comparable or slightly improved discriminatory capacity compared with 2D analyses, potentially reflecting more comprehensive tumor heterogeneity characterization. External validation cohorts were associated with slightly lower but more conservative performance estimates compared with internal validation studies, underscoring the importance of independent validation for clinical translation.

Combined models integrating clinical variables with radiomic features demonstrated improved diagnostic performance relative to imaging-only models, supporting the added value of multimodal risk stratification. Detailed pooled estimates for each subgroup are summarized in Table 3.

Table 3.

Subgroup Analysis of Pooled Diagnostic Performance of Radiomics-Based Models According to Modeling Method, Sample Size, Software Platform, Segmentation Strategy, Validation Approach, and Model Type.

Subgroup	Number of Studies	Sensitivity (95% CI)	I² Sens (%)	Specificity (95% CI)	I² Spec (%)	PLR	NLR	AUC
Modeling methods
Deep learning	14	0.75 (0.62-0.85)	75.6	0.82 (0.71-0.89)	90.7	4.17	0.30	0.847
Radiomics algorithm	20	0.79 (0.75-0.83)	0	0.76 (0.71-0.80)	61.2	3.29	0.28	0.857
Sample size
<100	14	0.71 (0.59-0.81)	50.9	0.78 (0.68-0.86)	83.8	3.23	0.37	0.820
≥100	14	0.81 (0.75-0.86)	51.3	0.79 (0.72-0.85)	87.6	3.86	0.24	0.878
Radiomic software
Others	20	0.77 (0.67-0.85)	59.0	0.79 (0.71-0.86)	84.7	3.67	0.29	0.862
PyRadiomics	14	0.78 (0.72-0.82)	28.8	0.74 (0.69-0.79)	70.7	3.00	0.30	0.841
Segmentation
3D	32	0.78 (0.73-0.82)	45.7	0.78 (0.72-0.82)	83.0	3.55	0.28	0.855
Validation
Split sample	15	0.76 (0.69-0.83)	50.1	0.77 (0.70-0.82)	72.4	3.30	0.31	0.851
External validation	19	0.79 (0.71-0.85)	46.0	0.79 (0.71-0.85)	86.9	3.76	0.27	0.855
Models
Combined model	32	0.78 (0.72-0.82)	52.3	0.77 (0.72-0.81)	82.4	3.39	0.29	0.849
Radiomics model	2	0.86 (0.42-0.98)	0	0.97 (0.90-0.99)	0	28.67	0.14	0.912

Note: Subgroup totals do not sum to the overall number of studies because some studies were not classifiable or did not report sufficient information for subgroup categorization. PLR and NLR values were calculated from pooled sensitivity and specificity estimates using standard diagnostic test accuracy formulas. Heterogeneity estimates for PLR and NLR were not reported because subgroup-level models were not fitted separately for these metrics.

Abbreviations: PLR, positive likelihood ratio; NLR, negative likelihood ratio; AUC, area under the receiver operating characteristic curve.

Publication Bias

Potential publication bias was assessed using Deeks’ funnel plot asymmetry test for diagnostic accuracy metaanalyses. Visual inspection of the funnel plot did not reveal marked asymmetry. Formal statistical testing using Deeks’ regression analysis demonstrated evidence of potential publication bias (P = .0009). The corresponding funnel plot is presented in Figure 6. These findings suggest the presence of potential publication bias and small-study effects, which may have led to overestimation of pooled diagnostic performance.

Figure 6.

Deeks’ funnel plot for publication bias. The funnel plot shows log diagnostic odds ratio (logDOR) against 1/√(ESS). The red regression line represents Deeks’ test for asymmetry. A significant slope (P = .0009) suggests potential publication bias.

Discussion

This systematic review and metaanalysis evaluated the diagnostic performance of MRI-based radiomics models for predicting pathological complete response (pCR) after neoadjuvant chemoradiotherapy in patients with locally advanced rectal adenocarcinoma. Based on 38 included studies and analysis of validation cohorts only, the pooled sensitivity was 0.82, and the pooled specificity was 0.86, with an overall AUC of 0.846. These findings indicate excellent discriminatory capacity and suggest that radiomics-based models may provide clinically meaningful noninvasive biomarkers for treatment response assessment.⁶¹

Accurate preoperative identification of pCR is of major clinical importance. Patients achieving pCR after neoadjuvant chemoradiotherapy demonstrate favorable long-term outcomes and may be candidates for organ-preserving strategies, including nonoperative “watch-and-wait” management. Avoidance of total mesorectal excision may substantially reduce morbidity related to bowel, urinary, and sexual dysfunction, as well as permanent stoma formation. Therefore, reliable imaging-based prediction tools could facilitate personalized therapeutic decision-making and optimize risk–benefit assessment.⁶²

Radiomic features capture intratumoral heterogeneity, which has been associated with treatment resistance and adverse oncologic outcomes.⁶³ Tumors failing to achieve pCR frequently demonstrate increased imaging heterogeneity, potentially reflecting residual viable tumor and microenvironmental complexity.⁶⁴ Similar associations between imaging heterogeneity and poor therapeutic response have been reported across solid tumors, supporting the biological plausibility of radiomics-based response prediction.⁶⁵

Our subgroup analyses suggested that DL models showed numerically comparable or slightly higher performance; however, no formal statistical comparison was conducted, and therefore these findings should be interpreted cautiously. DL architectures enable automated hierarchical feature learning directly from raw imaging data and can capture complex nonlinear relationships beyond handcrafted descriptors.⁶⁶ However, the number of DL studies remains limited, and many lack external validation.⁶⁷ Larger multicenter investigations are required to confirm these findings. The interpretation of subgroup analyses has been revised to reflect that observed differences between modeling approaches are descriptive and should be interpreted cautiously in the absence of formal statistical comparison.

Models integrating radiomic features with clinical variables demonstrated improved diagnostic performance compared with imaging-only models.⁶⁸ This observation aligns with the principle that tumor response is influenced by multiple biological and clinical factors. Multivariable predictive models combining imaging and nonimaging data may therefore provide superior risk stratification.

Despite encouraging diagnostic accuracy, methodological limitations were evident. Scanner variability and protocol heterogeneity may affect feature reproducibility, particularly in multicenter settings.⁶⁹ External validation remains inconsistently implemented in radiomics research.⁷⁰ Risk of bias assessment using QUADAS-2 identified concerns primarily in the index test domain. In many studies, model thresholds were not prespecified, increasing the potential for optimistic performance estimates.⁷¹ These findings underscore the importance of adherence to standardized reporting guidelines such as STARD.⁷² Although a formal quantitative sensitivity analysis stratified by risk of bias was not performed due to limited reporting consistency, studies judged as having lower risk of bias—particularly those with external validation and clearer threshold specification—tended to report more conservative performance estimates. This suggests that inclusion of studies with unclear or high risk of bias may have contributed to overestimation of pooled diagnostic accuracy. The presence of significant publication bias indicates that studies with favorable results may be overrepresented, potentially inflating pooled estimates of sensitivity and specificity.

Substantial heterogeneity was observed in pooled sensitivity and specificity. Heterogeneity is common in diagnostic test accuracy metaanalyses and may arise from differences in imaging protocols, segmentation strategies, feature selection procedures, and validation design.⁷³ Although subgroup analyses identified potential contributors, residual heterogeneity remained, highlighting the need for workflow standardization before clinical implementation. In addition to statistical heterogeneity, several methodological and clinical factors likely contributed to variability in diagnostic performance. Differences in MRI acquisition protocols, including field strength, sequence selection, and timing of posttreatment imaging, may influence radiomic feature stability and reproducibility. Furthermore, segmentation strategies (manual vs semi-automatic; 2D vs 3D) introduce variability in feature extraction, potentially affecting model performance. Variations in pCR definitions, although generally based on ypT0N0, may also contribute to inconsistencies across studies. These factors collectively limit comparability and highlight the need for standardized radiomics pipelines.

Emerging approaches such as delta radiomics, which quantify temporal changes in imaging features during treatment, may further enhance response prediction.⁷⁴ By capturing dynamic tumor alterations, delta models may provide additional biological insight beyond single-time-point imaging.

This study has several limitations. First, substantial heterogeneity persisted despite subgroup exploration. Second, although only validation cohorts were included to reduce overfitting bias, external validation remains limited. Third, only pCR was evaluated as the outcome of interest. In conclusion, MRI-based radiomics demonstrates high diagnostic performance for predicting pathological complete response after neoadjuvant chemoradiotherapy in rectal adenocarcinoma. However, methodological heterogeneity and limited prospective validation currently constrain clinical translation. Future standardized, multicenter, prospectively validated studies are essential to establish radiomics as a reliable decision-support tool in personalized rectal cancer management. The exclusion of non-English studies may have introduced language bias, potentially limiting the generalizability of the findings.

Future studies should incorporate standardized imaging protocols, multicenter external validation cohorts with adequate sample sizes (preferably >200 patients), adherence to Image Biomarker Standardization Initiative (IBSI) guidelines, transparent model reporting, and prospective study designs to ensure clinical applicability.

Conclusions

This systematic review and metaanalysis demonstrates that MRI-based radiomics provides promising diagnostic accuracy but is not yet ready for routine clinical implementation. The pooled results indicate that radiomics has substantial potential as a noninvasive imaging biomarker to support individualized treatment strategies. Subgroup analyses suggest that DL-based models may achieve superior predictive performance compared with conventional radiomics-based ML approaches, and that integration of clinical variables further enhances diagnostic accuracy. These findings highlight the value of multimodal and data-driven predictive frameworks in rectal cancer management. However, methodological heterogeneity, limited prospective validation, and incomplete adherence to radiomics quality standards currently restrict routine clinical implementation.

Supplemental Material

sj-doc-1-mix-10.1177_15353508261462778 - Supplemental material for Radiomic Analysis of MRI for Assessing Response to Neoadjuvant Chemoradiotherapy in Rectal Adenocarcinoma: A Systematic Review and Metaanalysis

Supplemental material, sj-doc-1-mix-10.1177_15353508261462778 for Radiomic Analysis of MRI for Assessing Response to Neoadjuvant Chemoradiotherapy in Rectal Adenocarcinoma: A Systematic Review and Metaanalysis by Murat Jakipov, MD, MSc, PhD, Amin Tamadon, PhD, Zhandos Burkitbayev, MD, PhD, Bayram Kochiev, MD, Aslan Karimov, MD, Aigerim Temirbayeva, MD, Yerbolat Iztleuov, MD, PhD, Prashant Jamwal, PhD, Keivan Daneshvar, MD, Nadiar M. Mussin, MD, PhD and Ramazon Safarzoda Sharoffidin, PhD in SAGE Publications

Footnotes

Acknowledgments

Not applicable.

ORCID iDs

Amin Tamadon

Bayram Kochiev

Aslan Karimov

Ramazon Safarzoda Sharoffidin

Institutional Review Board Statement

Not applicable.

Informed Consent

Not applicable.

Author Contributions

MJ, AT, and RSS: conceptualization; MJ, AT, PJ, and KD: methodology; PJ, KD, and RSS: formal analysis; MJ, ZB, BK, AK, and AT: investigation; MJ, ZB, BK, AK, AT, YI: data curation; AT and YI: validation; AT and MJ: writing–original draft; PJ, ZB, BK, AK, KD, YI, NMM, and RSS: writing–review and editing; YI and AT: supervision; AT: project administration.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The datasets generated and analyzed during the current study are derived from publicly available studies included in this systematic review and metaanalysis. All relevant data supporting the findings of this study are included within the article and its Supplemental materials. Additional extracted data used for the metaanalysis are available from the corresponding author upon reasonable request.

Supplemental Material

Supplemental material for this article is available online.

References

Sung

Ferlay

Siegel

, et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71(3):209‐249. doi:10.3322/caac.21660

Sauer

Becker

Hohenberger

, et al. German Rectal Cancer Study G. Preoperative versus postoperative chemoradiotherapy for rectal cancer. N Engl J Med. 2004;351(17):1731‐1740. doi:10.1056/NEJMoa040694

Maas

Nelemans

Valentini

, et al. Long-term outcome in patients with a pathological complete response after chemoradiation for rectal cancer: A pooled analysis of individual patient data. Lancet Oncol. 2010;11(9):835‐844. doi:10.1016/S1470-2045(10)70172-8

Fokas

Liersch

Fietkau

, et al. Tumor regression grading after preoperative chemoradiotherapy for locally advanced rectal carcinoma revisited: Updated results of the CAO/ARO/AIO-94 trial. J Clin Oncol. 2014;32(15):1554‐1562. doi:10.1200/JCO.2013.54.3769

Zorcolo

Rosman

Restivo

, et al. Complete pathologic response after combined modality treatment for rectal cancer and long-term survival: A meta-analysis. Ann Surg Oncol. 2012;19(9):2822‐2832. doi:10.1245/s10434-011-2209-y

van der Valk

MJM

Hilling

Bastiaannet

, et al. Long-term outcomes of clinical complete responders after neoadjuvant treatment for rectal cancer in the International Watch & Wait Database (IWWD): An international multicentre registry study. Lancet. 2018;391(10139):2537‐2545. doi:10.1016/S0140-6736(18)31078-X

Bryant

Lunniss

Knowles

Thaha

Chan

. Anterior resection syndrome. Lancet Oncol. 2012;13(9):e403‐e408. doi:10.1016/S1470-2045(12)70236-X

Habr-Gama

Perez

Nadalin

, et al. Operative versus nonoperative treatment for stage 0 distal rectal cancer following chemoradiation therapy: Long-term results. Ann Surg. 2004;240(4):711‐717; discussion 7-8. doi:10.1097/01.sla.0000141194.27992.32

Beets-Tan

. MRI in rectal cancer: The T stage and circumferential resection margin. Colorectal Dis. 2003;5(5):392‐395. doi:10.1046/j.1463-1318.2003.00518.x

10.

Patel

Taylor

Blomqvist

, et al. Magnetic resonance imaging-detected tumor response for locally advanced rectal cancer predicts survival outcomes: MERCURY experience. J Clin Oncol. 2011;29(28):3753‐3760. doi:10.1200/JCO.2011.34.9068

11.

Lambregts

Cappendijk

Maas

Beets

Beets-Tan

. Value of MRI and diffusion-weighted MRI for the diagnosis of locally recurrent rectal cancer. Eur Radiol. 2011;21(6):1250‐1258. doi:10.1007/s00330-010-2052-8

12.

, et al. A noninvasive tool based on magnetic resonance imaging radiomics for the preoperative prediction of pathological complete response to neoadjuvant chemotherapy in breast cancer. Ann Surg Oncol. 2022;29(12):7685‐7693. doi:10.1245/s10434-022-12034-w

13.

Scapicchio

Gabelloni

Barucci

Cioni

Saba

Neri

. A deep look into radiomics. Radiol Med. 2021;126(10):1296‐1311. doi:10.1007/s11547-021-01389-x

14.

Ganeshan

Miles

. Quantifying tumour heterogeneity with CT. Cancer Imaging. 2013;13(1):140‐149. doi:10.1102/1470-7330.2013.0015

15.

Feng

Wang

Jiao

. Integrating radiomics and machine learning for the diagnosis and prognosis of hepatocellular carcinoma. World J Gastrointest Oncol. 2025;17(7):106610. doi:10.4251/wjgo.v17.i7.106610

16.

Huang

Lin

Deng

Tang

. Radiomics in rectal cancer: Current status of use and advances in research. Front Oncol. 2024;14:1470824. doi:10.3389/fonc.2024.1470824

17.

Wang

Liu

. Machine learning in predicting pathological complete response to neoadjuvant chemoradiotherapy in rectal cancer using MRI: A systematic review and meta-analysis. Br J Radiol. 2024;97(1159):1243‐1254. doi:10.1093/bjr/tqae098

18.

Nouroozi

Kazemi

Alinezhad

, et al. Artificial intelligence-based detection of neuropsychiatric lupus: An exploratory meta-analysis of neuroimaging and multimodal biomarker models. Clin Exp Med. 2026;26(1):125. doi:10.1007/s10238-025-02030-1

19.

Flaiban

Orhan

Goncalves

Lopes

Costa

ALF

. Radiomics in action: Multimodal synergies for imaging biomarkers. Bioengineering (Basel). 2025;12(11):1139. doi:10.3390/bioengineering12111139

20.

Mese

Kocak

. Evaluating methodological quality in radiomics research using large language models: Added value of METRICS-E3 framework. Eur J Radiol. 2026;194:112519. doi:10.1016/j.ejrad.2025.112519

21.

Whiting

Rutjes

Westwood

, et al. QUADAS-2: A revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. 2011;155(8):529‐536. doi:10.7326/0003-4819-155-8-201110180-00009

22.

Lambin

Leijenaar

RTH

Deist

, et al. Radiomics: The bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749‐762. doi:10.1038/nrclinonc.2017.141

23.

Alvarez-Jimenez

Antunes

DeSilvio

, et al. A novel structural modeling magnitude and orientation radiomic descriptor for evaluating response to neoadjuvant therapy in rectal cancers via MRI. NPJ Precis Oncol. 2025;9(1):215. doi:10.1038/s41698-025-01007-3

24.

Antunes

Ofshteyn

Bera

, et al. Radiomic features of primary rectal cancers on baseline T(2)-weighted MRI are associated with pathologic complete response to neoadjuvant chemoradiation: A multisite study. J Magn Reson Imaging. 2020;52(5):1531‐1541. doi:10.1002/jmri.27140

25.

Azamat

Karaman

Azamat

, et al. Complete response evaluation of locally advanced rectal cancer to neoadjuvant chemoradiotherapy using textural features obtained from T2 weighted imaging and ADC maps. Curr Med Imaging. 2022;18(10):1061‐1069. doi:10.2174/1573405618666220303111026

26.

Begal

Sabo

Goldberg

Bitterman

Khoury

. Wavelets-based texture analysis of post neoadjuvant chemoradiotherapy magnetic resonance imaging as a tool for recognition of pathological complete response in rectal cancer, a retrospective study. J Clin Med. 2024;13(23):7383. doi:10.3390/jcm13237383

27.

Bellini

Carbone

Rengo

, et al. Performance of machine learning and texture analysis for predicting response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer with 3T MRI. Tomography. 2022;8(4):2059‐2072. doi:10.3390/tomography8040173

28.

Boldrini

Lenkowicz

Orlandini

, et al. Applicability of a pathological complete response magnetic resonance-based radiomics model for locally advanced rectal cancer in intercontinental cohort. Radiat Oncol. 2022;17(1):78. doi:10.1186/s13014-022-02048-9

29.

Bulens

Couwenberg

Intven

, et al. Predicting the tumor response to chemoradiotherapy for rectal cancer: Model development and external validation using MRI radiomics. Radiother Oncol. 2020;142:246‐252. doi:10.1016/j.radonc.2019.07.033

30.

Chen

Xie

, et al. MRI-based Radiomics features to predict treatment response to neoadjuvant chemotherapy in locally advanced rectal cancer: A single center, prospective study. Front Oncol. 2022;12:801743. doi:10.3389/fonc.2022.801743

31.

Cheng

Luo

, et al. Multiparametric MRI-based radiomics approaches on predicting response to neoadjuvant chemoradiotherapy (nCRT) in patients with rectal cancer. Abdom Radiol (NY). 2021;46(11):5072‐5085. doi:10.1007/s00261-021-03219-0

32.

Crimi

D'Alessandro

Zanon

, et al. A machine learning model based on MRI radiomics to predict response to chemoradiation among patients with rectal cancer. Life (Basel). 2024;14(12):1530. doi:10.3390/life14121530

33.

Cui

Yang

Shi

, et al. Radiomics analysis of multiparametric MRI for prediction of pathological complete response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Eur Radiol. 2019;29(3):1211‐1220. doi:10.1007/s00330-018-5683-9

34.

Feng

Liu

, et al. Development and validation of a radiopathomics model to predict pathological complete response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer: A multicentre observational study. Lancet Digit Health. 2022;4(1):e8‐e17. doi:10.1016/S2589-7500(21)00215-6

35.

Horvat

Veeraraghavan

Khan

, et al. MR imaging of rectal cancer: Radiomics analysis to assess treatment response after neoadjuvant therapy. Radiology. 2018;287(3):833‐843. doi:10.1148/radiol.2018172300

36.

Horvat

Veeraraghavan

Nahas

CSR

, et al. Combined artificial intelligence and radiologist model for predicting rectal cancer treatment response from magnetic resonance imaging: An external validation study. Abdom Radiol (NY). 2022;47(8):2770‐2782. doi:10.1007/s00261-022-03572-8

37.

Gong

Sun

, et al. Magnetic resonance imaging-based radiomics analysis for prediction of treatment response to neoadjuvant chemoradiotherapy and clinical outcome in patients with locally advanced rectal cancer: A large multicentric and validated study. MedComm. (2020). 2024;5(7):e609. doi:10.1002/mco2.609

38.

Huang

Han

Guo

, et al. Multiphase and multiparameter MRI-based radiomics for prediction of tumor response to neoadjuvant therapy in locally advanced rectal cancer. Radiat Oncol. 2023;18(1):179. doi:10.1186/s13014-023-02368-4

39.

Jang

Lim

Song

, et al. Image-based deep learning model for predicting pathological response in rectal cancer using post-chemoradiotherapy magnetic resonance imaging. Radiother Oncol. 2021;161:183‐190. doi:10.1016/j.radonc.2021.06.019

40.

Jiang

Guo

, et al. A comprehensive prediction model based on MRI radiomics and clinical factors to predict tumor response after neoadjuvant chemoradiotherapy in rectal cancer. Acad Radiol. 2023;30(Suppl 1):S185‐SS98. doi:10.1016/j.acra.2023.04.032

41.

Jin

, et al. Predicting treatment response from longitudinal images using multi-task deep learning. Nat Commun. 2021;12(1):1851. doi:10.1038/s41467-021-22188-y

42.

Lee

Lim

Shin

Kim

Hwang

. Pathologic complete response prediction after neoadjuvant chemoradiation therapy for rectal cancer using radiomics and deep embedding network of MRI. Appl Sci. 2021;11(20):9494. doi:10.3390/app11209494

43.

Lee

Kim

Seo

Moon

Park

You

. Machine learning-based response assessment in patients with rectal cancer after neoadjuvant chemoradiotherapy: Radiomics analysis for assessing tumor regression grade using T2-weighted magnetic resonance images. Int J Colorectal Dis. 2024;39(1):78. doi:10.1007/s00384-024-04651-6

44.

Yuan

Liu

, et al. Predicting pathological complete response following neoadjuvant chemoradiotherapy (nCRT) in patients with locally advanced rectal cancer using merged model integrating MRI-based radiomics and deep learning data. BMC Med Imaging. 2024;24(1):289. doi:10.1186/s12880-024-01474-3

45.

Miranda

Horvat

Assuncao

Jr ., et al. MRI-based radiomic score increased mrTRG accuracy in predicting rectal cancer response to neoadjuvant therapy. Abdom Radiol (NY). 2023;48(6):1911‐1920. doi:10.1007/s00261-023-03898-x

46.

Nardone

Reginelli

Grassi

, et al. Ability of delta radiomics to predict a complete pathological response in patients with loco-regional rectal cancer addressed to neoadjuvant chemo-radiation and surgery. Cancers (Basel). 2022;14(12):3004. doi:10.3390/cancers14123004

47.

Pang

Wang

Zhang

, et al. A pipeline for predicting the treatment response of neoadjuvant chemoradiotherapy for locally advanced rectal cancer using single MRI modality: Combining deep segmentation network and radiomics analysis based on “suspicious region”. Front Oncol. 2021;11:711747. doi:10.3389/fonc.2021.711747

48.

Peng

Wan

Wang

Zou

Zhao

Zhang

. A multiple-time-scale comparative study for the added value of magnetic resonance imaging-based radiomics in predicting pathological complete response after neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Front Oncol. 2023;13:1234619. doi:10.3389/fonc.2023.1234619

49.

Rengo

Landolfi

Picchia

, et al. Rectal cancer response to neoadjuvant chemoradiotherapy evaluated with MRI: Development and validation of a classification algorithm. Eur J Radiol. 2022;147:110146. doi:10.1016/j.ejrad.2021.110146

50.

Shaish

Aukerman

Vanguri

, et al. Radiomics of MRI for pretreatment prediction of pathologic complete response, tumor regression grade, and neoadjuvant rectal score in patients with locally advanced rectal cancer undergoing neoadjuvant chemoradiation: An international multicenter study. Eur Radiol. 2020;30(11):6263‐6273. doi:10.1007/s00330-020-06968-6

51.

Shin

Seo

Baek

, et al. MRI radiomics model predicts pathologic complete response of rectal cancer following chemoradiotherapy. Radiology. 2022;303(2):351‐358. doi:10.1148/radiol.211986

52.

Shen

, et al. Combining clinicopathology, IVIM-DWI and texture parameters for a nomogram to predict treatment response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer patients. Front Oncol. 2022;12:886101. doi:10.3389/fonc.2022.886101

53.

Wan

Zhang

Zhao

, et al. Developing a prediction model based on MRI for pathological complete response after neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Abdom Radiol (NY). 2019;44(9):2978‐2987. doi:10.1007/s00261-019-02129-6

54.

Wan

Peng

Zou

, et al. MRI-based delta-radiomics are predictive of pathological complete response after neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Acad Radiol. 2021;28(Suppl 1):S95‐S104. doi:10.1016/j.acra.2020.10.026

55.

Wang

Zhang

Jiang

, et al. Multiparametric magnetic resonance imaging (MRI)-based radiomics model explained by the Shapley additive exPlanations (SHAP) method for predicting complete response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer: A multicenter retrospective study. Quant Imaging Med Surg. 2024;14(7):4617‐4634. doi:10.21037/qims-24-7

56.

Wen

Liu

, et al. MRI-based radiomic models outperform radiologists in predicting pathological complete response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Acad Radiol. 2023;30(Suppl 1):S176‐SS84. doi:10.1016/j.acra.2022.12.037

57.

Yardimci

Kocak

Sel

, et al. Radiomics of locally advanced rectal cancer: Machine learning-based prediction of response to neoadjuvant chemoradiotherapy using pre-treatment sagittal T2-weighted MRI. Jpn J Radiol. 2023;41(1):71‐82. doi:10.1007/s11604-022-01325-7

58.

Pei

Zhang

, et al. MRI-based radiomics predicts tumor response to neoadjuvant chemoradiotherapy in locally advanced rectal cancer. Front Oncol. 2019;9:552. doi:10.3389/fonc.2019.00552

59.

Zhang

Wang

Zhu

, et al. Predicting rectal cancer response to neoadjuvant chemoradiotherapy using deep learning of diffusion kurtosis MRI. Radiology. 2020;296(1):56‐64. doi:10.1148/radiol.2020190936

60.

Zhu

Zhang

Shi

Sun

. The conversion of MRI data with multiple b-values into signature-like pictures to predict treatment response for rectal cancer. J Magn Reson Imaging. 2022;56(2):562‐569. doi:10.1002/jmri.28033

61.

Agosti

Mapelli

Grimod

Piazza

Fontanella

Panciani

. MRI-based radiomics for non-invasive prediction of molecular biomarkers in gliomas. Cancers (Basel). 2026;18(3):491. doi:10.3390/cancers18030491

62.

Cusumano

Dinapoli

Boldrini

, et al. Fractal-based radiomic approach to predict complete pathological response after chemo-radiotherapy in rectal cancer. Radiol Med. 2018;123(4):286‐295. doi:10.1007/s11547-017-0838-3

63.

Henry

Sun

Lerousseau

, et al. Investigation of radiomics based intra-patient inter-tumor heterogeneity and the impact of tumor subsampling strategies. Sci Rep. 2022;12(1):17244. doi:10.1038/s41598-022-20931-z

64.

Janiszewska

Stein

Metzger Filho

, et al. The impact of tumor epithelial and microenvironmental heterogeneity on treatment responses in HER2+breast cancer. JCI Insight. 2021;6(11):e147617. doi:10.1172/jci.insight.147617

65.

Yip

Aerts

. Applications and limitations of radiomics. Phys Med Biol. 2016;61(13):R150‐R166. doi:10.1088/0031-9155/61/13/R150

66.

Trebeschi

van Griethuysen

JJM

Lambregts

DMJ

, et al. Deep learning for fully-automated localization and segmentation of rectal cancer on multiparametric MR. Sci Rep. 2017;7(1):5301. doi:10.1038/s41598-017-05728-9

67.

Yuan

Gan

, et al. A comparative study between deep learning and radiomics models in grading liver tumors using hepatobiliary phase contrast-enhanced MR images. BMC Med Imaging. 2022;22(1):218. doi:10.1186/s12880-022-00946-8

68.

Ching

JCF

Lam

CCH

, et al. Integrating CT-based radiomic model with clinical features improves long-term prognostication in high-risk prostate cancer. Front Oncol. 2023;13:1060687. doi:10.3389/fonc.2023.1060687

69.

Demircioglu

. Reproducibility and interpretability in radiomics: A critical assessment. Diagn Interv Radiol. 2025;31(4):321‐328. doi:10.4274/dir.2024.242719

70.

Hagiwara

Fujita

Ohno

Aoki

. Variability and standardization of quantitative imaging: Monoparametric to multiparametric quantification, radiomics, and artificial intelligence. Invest Radiol. 2020;55(9):601‐616. doi:10.1097/RLI.0000000000000666

71.

Leucuta

Urda-Cimpean

Istrate

Drugan

. Risk of bias assessment of diagnostic accuracy studies using QUADAS 2 by large language models. Diagnostics (Basel). 2025;15(12):1451. doi:10.3390/diagnostics15121451

72.

Cohen

Korevaar

Altman

, et al. STARD 2015 guidelines for reporting diagnostic accuracy studies: Explanation and elaboration. BMJ Open. 2016;6(11):e012799. doi:10.1136/bmjopen-2016-012799

73.

White

Phua

Yaxley

McInnes

MDF

. Heterogeneity in systematic reviews of medical imaging diagnostic test accuracy studies: A systematic review. JAMA Netw Open. 2024;7(2):e240649. doi:10.1001/jamanetworkopen.2024.0649

74.

Boldrini

Cusumano

Chiloiro

, et al. Delta radiomics for rectal cancer response prediction with hybrid 0.35 T magnetic resonance-guided radiotherapy (MRgRT): A hypothesis-generating study for an innovative personalized medicine approach. Radiol Med. 2019;124(2):145‐153. doi:10.1007/s11547-018-0951-y

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.03 MB

0.00 MB