Key Technology Considerations in Developing and Deploying Machine Learning Models in Clinical Radiology Practice

Abstract

The use of machine learning to develop intelligent software tools for the interpretation of radiology images has gained widespread attention in recent years. The development, deployment, and eventual adoption of these models in clinical practice, however, remains fraught with challenges. In this paper, we propose a list of key considerations that machine learning researchers must recognize and address to make their models accurate, robust, and usable in practice. We discuss insufficient training data, decentralized data sets, high cost of annotations, ambiguous ground truth, imbalance in class representation, asymmetric misclassification costs, relevant performance metrics, generalization of models to unseen data sets, model decay, adversarial attacks, explainability, fairness and bias, and clinical validation. We describe each consideration and identify the techniques used to address it. Although these techniques have been discussed in prior research, by freshly examining them in the context of medical imaging and compiling them in the form of a laundry list, we hope to make them more accessible to researchers, software developers, radiologists, and other stakeholders.

A Classical-Quantum Convolutional Neural Network for Detecting Pneumonia from Chest Radiographs

“…The performance of a machine learning system in practice depends on a large number of considerations [47]. State-of-art neural network architectures for image classification contain millions of trainable parameters, and they are trained on hundreds of thousands of scans [37].…”

Levels of Autonomous Radiology

“…Specialized explainer systems of explainable AI (widely acknowledged as an important feature of practical deployment of AI models) aim at explaining AI inferences to human users [ 19 ]. Explainability in radiology can be improved by using localization models, which can highlight the region of suspected abnormality (region of interest) in the scan, instead of using classification models, which only indicate the presence or absence of an abnormality [ 20 ]. Although an explainable system does not refer to an explicit human model and only indicates or highlights the decision-relevant parts of the AI model (ie, parts that contributed to a specific prediction), causability refers primarily to a human-understandable model.…”

“…This may lead to uncertainty in the ground truth labels. The problem of ambiguous ground truth can be mitigated by using expert adjudication [ 21 ] or multiphasic review [ 22 ] to create high-quality labels, which may help yield better models than other approaches in improving model performance on original labels [ 20 ]. Additionally, imaging protocols, manufacturers of imaging modalities, and the process of storing and processing medical data differ between organizations, which impedes the use of data from different sources for AI applications [ 23 ].…”

“…Such systems would require extensive clinical validations before they could be reliably used. As summarized by Kulkarni et al [ 20 ] , these systems would need to undergo internal as well as external validations to evaluate the system’s performance on unseen data. The Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement [ 25 ] specifies guidelines for reporting the development and validation of such diagnostic or prognostic models.…”

“…These changes may occur gradually over time or suddenly when the AI system is deployed in a different location with a different population. Therefore, it is crucial to continuously monitor these AI systems and fine-tune them as required to maintain optimal performance [ 20 ].…”