Best of Both Worlds: Deep Learning Reconstruction Reduces MRI Acquisition Time and Improves Image Quality

Abstract

Magnetic resonance imaging (MRI) is commonly used in the evaluation of pancreatic and hepatobiliary pathologies because it offers superior delineation of the pancreatic and biliary ductal anatomy and superior tissue contrast compared to computed tomography. MRI also does not require ionizing radiation, which is a significant advantage for patients requiring serial exams. Imaging findings of pancreatic and hepatobiliary diseases can be subtle, including small lesions, minor textural change and intensity differences, mural nodularity, and ductal irregularity. Detection of these subtle changes critically depends on MRI image quality, and poor-quality images due to motion artifact, poor signal to noise (SNR), and poor contrast to noise (CNR), can significantly reduce our diagnostic accuracy. Reduced field-of-view (rFOV) MR sequence can improve the in-plane spatial resolution and can potentially improve the delineation of fine anatomic detail and increase the lesion conspicuity. However, this focused approach is associated with increased image noise and scan duration.

Deep learning reconstruction (DLR) offers a potential solution for accelerating MRI acquisition while maintaining or enhancing image quality. Faster MRI acquisitions generally lead to noisier or lower-spatial resolution images, and DLR can be used to reduce the image noise or improve spatial resolution.^1,2 This can be accomplished at the sensor space (k-space), image space, or with a hybrid approach. k-space acceleration tool is usually directly installed on the MRI scanner and may be available only on newer scanners or software platforms. Image-space acceleration is applied after the reconstruction process, and the unenhanced images can be theoretically reviewed along with the enhanced images, which may increase reader confidence in the fidelity of the enhanced images.¹

T2-weighted (T2WI) and diffusion weighted imaging (DWI) have the longest acquisition times among abdominal MRI sequences, and they are the prime candidates for DLR MR acceleration. Recent review article by Rajamohan et al showed that DLR acceleration could achieve scan time reduction between 21% and 93% for T2WI and between 24% and 62% for DWI in comparison with conventional approaches.² Beyond significant savings in image acquisition time, DLR has been shown to improve SNR, CNR, and lesion conspicuity, which highlighted the potential for DLR in improving image quality and workflow efficiency.

In this issue, Wang et al evaluated the impact of DLR of rFOV T2WI of pancreaticobiliary disorders.³ This prospective study included 198 patients who underwent respiratory triggered rFOV T2WI obtained on 3 Tesla MRI scanners. The protocol included the standard rFOV T2WI sequence with number of excitations (NEX) of 2 (rFOV T2WI_N2), and a shorter sequence with NEX of 1 (rFOV T2WI_N1). After the data was acquired, a vendor specific DLR algorithm was applied to both sequences (NEX = 1 and NEX = 2). Qualitative and quantitative assessment of image quality was performed. Three observers performed the qualitative analysis in a blinded fashion to determine noise, respiratory motion artifacts, overall image quality, and diagnostic confidence. Two independent observers measured the SNR and the CNR for the pancreas.

The participants in this study had various underlying pancreatic and hepatobiliary pathologies, including solid pancreatic masses, cystic pancreatic masses, pancreatitis, cholangiocarcinoma, gallbladder carcinoma, cholelithiasis, and choledocholithiasis. The rFOV T2WI_N1 sequence acquisition time was significantly shorter compared to the rFOV T2WI_N2 sequence (109 ± 22.0 seconds vs 233.2 ± 46.6 seconds), which translated to 37.1% to 74.4% improvement in efficiency. However, the tradeoff of the shorter acquisition (NEX = 1) resulted in inferior subjective image quality, SNR, and CNR. Importantly, the shorter acquisition also reduced the detection rate of pancreaticobiliary lesions from 91.0% to 76.5% (P < .001). This highlighted and validated concerns with traditional MRI acceleration: reduction in MRI acquisition time may degrade image quality and lead to loss of diagnostic information.

Then the authors showed that the application of DLR to the rFOV T2WI_N1 sequence could effectively compensate for the loss of image quality. DLR improved the subjective image quality, SNR, and CNR, and achieved statistically equivalent performance to the standard sequence rFOV T2WI_N2. In fact, there was reduced respiratory motion in the rFOV T2WI_N1-DLR sequence compared to the rFOV T2WI_N2, due to the shorter acquisition. Applying DLR to the shorter sequence (rFOV T2WI_N1) also improved the pancreatobiliary lesion detection rate from 76.5% to 92.7% and recovered important diagnostic information that was lost with the shorter acquisition. The results of this study validated results of previous studies^4-6 which suggested that DLR MRI acceleration could decrease image acquisition time without sacrificing image quality.

This study had several limitations. This study was performed with a specific vendor-implemented DLR on a single 3.0 Tesla system at a single institution. These promising results should be validated in future studies on scanners from other vendors and at 1.5 Tesla. In future studies, DLR could also be used in conjunction with other commonly used acceleration techniques such as parallel imaging or compressed sensing to further decrease acquisition time and/or enhance image quality. Development of vendor neutral platforms may also promote broader dissemination of this technology. While the future of DLR in MRI acceleration seems bright, there are some pitfalls that warrant our attention. DLR has the potential for hallucination, invention of lesions that do not exist or removal of true lesions.¹ Some of these invented lesions can be easily recognized by the radiologists as artifactual, and do not pose any clinical concern. However, other artifacts or hallucinations may be subtle. Radiologists should carefully scrutinize the DLR images and may suspect hallucinations if they detect inconsistencies on different image sections and image types. Fortunately, current evidence suggests that hallucinations are infrequent for typical acceleration factors (2-4×) and may not affect diagnostic performance.¹ Furthermore, DLR may cause minor changes in the appearance of the images compared to standard sequences. This may have a negative impact on the performance of post-processing tools that were developed on non-DLR sequences.¹

In summary, DLR has the potential to reduce acquisition time and improve image quality in patients undergoing T2WI for pancreaticobiliary disorders. This improved workflow efficiency can also enhance patient access and operational efficiency for radiology departments.

Footnotes

ORCID iD

Linda C. Chu

Funding

The author disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Dr. Linda C. Chu receives research salary support from The Lustgarten Foundation.

References

Haller

Hedderich

Federau

, et al. The current status of AI-accelerated MRI techniques in clinical use. Radiology. 2025;317(2):e243819.

Rajamohan

Bagga

Bansal

Ginocchio

Gupta

Chandarana

Deep learning-accelerated MRI in body and chest. J Comput Assist Tomogr. 2025;49(4):531-544.

Wang

Liu

Wang

, et al. Accelerated reduced field of view T2-weighted imaging of pancreatobiliary disorders using deep learning-based reconstruction: reduction of acquisition time and improvement of image quality. Can Assoc Radiol J. Published online 2026. doi: 10.1177/08465371251407889.

Wary

Hossu

Ambarki

, et al. Deep learning HASTE sequence compared with T2-weighted BLADE sequence for liver MRI at 3 Tesla: a qualitative and quantitative prospective study. Eur Radiol. 2023;33(10):6817-6827.

Mulé

Kharrat

Zerbib

, et al. Fast T2-weighted liver MRI: image quality and solid focal lesions conspicuity using a deep learning accelerated single breath-hold HASTE fat-suppressed sequence. Diagn Interv Imaging. 2022;103(10):479-485.

Zhou

Qin

, et al. Thin-slice T₂-weighted images and deep-learning-based super-resolution reconstruction: improved preoperative assessment of vascular invasion for pancreatic ductal adenocarcinoma. Insights Imaging. 2025;16(1):144.