|
Recent AI breast cancer risk prediction models are difficult to interpret, limiting their clinical utility. In this work we explore the explainability of an AI-based risk-prediction model by examining performance with respect to different characteristics of the future cancer. In particular, saliency maps were used to examine how often the model focused on regions coinciding with future lesions and assess the characteristics of future lesions that were most likely to coincide with AI-assigned high-risk regions. An AI model for breast cancer risk prediction was previously trained on the UK OPTIMAM dataset, achieving an AUROC of 0.70 for the task of 3-year risk prediction. Re-visiting the test set used to evaluate this model (n=31351 examinations), we obtained additional information about the future cancer cases (n=1053), including future cancer type (invasive/in-situ) and grade, and future lesion visual characteristics. Patient-level risk was compared across different cancer types and grades, and saliency maps were generated to perform a localisation study. The AI tool performed similarly for future invasive and in-situ disease, with no significant difference in risk score observed. Similarly, risk scores did not vary significantly with future cancer grade. Saliency map analysis showed that the AI-indicated high-risk regions coincided more often with the location of future obvious lesions or lesions with calcifications. The results in this work provide insights into the decision-making process of the AI risk prediction tool. Further work is required to explore additional lesion characteristics and further validate these findings.
|
