Using Counterfactual Information for Breast Classification Diagnosis

Miguel Cardoso, Carlos Santiago, Jacinto C. Nascimento; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 4996-5002

The last radiology report by the Royal College of Radiologists has identified the pressure that radiologists are suffering due to excessive workloads levels. This is due to the availability of a growing number of images and a short time to provide the report making the diagnosis a difficult process. This suggests that the visualization of the radiologist should be accompanied somehow by an automatic "explainable" process. In this paper we give emphasis to the breast cancer as it is one of the most common types of cancer in women. Specifically we deal with mammography images because it is a primary step to be accomplished in an early radiological breast diagnosis. Although machine learning models are being used in medical imaging these models still struggle to provide enough interpretability to provide reliability in the decision-making process of the radiologist. In this work we explore solutions that improve an explainable model's performance in mammography classification. We propose the use counterfactual information for improving the breast classification task. We compare multiple approaches to the integration of counterfactual information into the training process. The experimental evaluation testifies that incorporating such counterfactual information improves both balanced accuracy and interpretability for the breast classification task.