Explaining Eye Diseases Detected by Machine Learning Using SHAP: A Case Study of Diabetic Retinopathy and Choroidal Nevus

Esmaeil Shakeri, Trafford Crump, Ezekiel Weis, Emad Mohammed, Roberto Souza, Behrouz H. Far

Abstract: Most visual impairment and eye cancers are preventable if detected in their early stages. Diabetic retinopathy (DR) is a significant cause of blindness worldwide and a serious public health concern in a population aged 20–65. With the increasing number of diabetes globally and its effects on patients’ vision, the automatic detection of DR has received wide attention from the machine learning field. Uveal melanoma (UM) is one of the most severe intraocular cancers in adults aged 50–80. A choroidal nevus (CN) is one of the most common intraocular tumours that can transform into UM, which can cause eyesight loss and spiteful melanoma with a high risk of melanoma-relevant metastasis and even death. Early prediction of UM can mitigate the risk of death caused by skeptical diagnosis decisions. In this paper, we use a transfer learning technique with a convolutional neural network (CNN)-based algorithm to detect UM and improve the interpretation of the diagnosis results. However, due to the black-box nature of deep learning and machine learning models, the interpretation and reliability of the predictions are still an issue that needs to be addressed before deploying these predictive models successfully. In this paper, we use the SHapley Additive exPlanations (SHAP) analysis approach to detect areas of an eye image that contribute the most to the DR and CN prediction using transfer learning. Our predictive model achieves an accuracy of 97% and 81% for binary and multi-class classification of DR and 82.5% accuracy for binary classification of CN. The SHAP analysis of the proposed method shows that regardless of the performance of the predictive models, this approach can be used as a tool to interpret the prediction results with more context-sensitive information about each sample and better understand the reasons for the classification results.