Interpretable deep learning approach for classification of breast cancer - a comparative analysis of multiple instance learning models

Breast cancer is the most frequent female cancer. Its early diagnosis increases the chances of a complete cure for the patient. Suitably designed deep learning algorithms can be an excellent tool for quick screening analysis and support radiologists and oncologists in diagnosing breast cancer. The design of a deep learning-based system for automated breast cancer diagnosis is not easy due to the lack of annotated data, especially at pixel level, the large size of the images with relatively small cancer lesion sizes and class imbalance, a wide diversity of cancerous lesions, a variety of breasts, both in size and density, make the training of the neural models challenging. Moreover, clinicians are often concerned about using these black-box models because of the lack of transparency in their inference. To address these issues, we propose an approach taking advantage of Multiple Instance Learning (MIL), supported by attention mechanisms. We researched Attention-based MIL (AMIL), Gated AMIL (GAMIL), Dual Stream MIL (DSMIL) and CLustering-constrained AMIL (CLAM) models trained in a weakly-supervised manner and compared them with a common model in image classification tasks, ResNet18. The approach described in this paper is multimodal and combines two mammographic projections (CC and MLO) in the training process. The developed neural system achieved high classification efficiency. Furthermore, exploiting the generated attentional maps allowed the localisation of cancerous lesions, thus increasing the interpretability of the algorithm. Thanks to this mechanism, we were also able to detect artifacts in the analyzed database, difficult to spot but drastically skewing the algorithm's performance.