TransRAM: A classification model of breast pathology images that imitates the diagnosis process of pathologists
Weiming Mi, Tao Zhang, Jiaming Wang · 2021 China Automation Congress (CAC) · 2021
Breast cancer, one of the most common health threats to females worldwide, has always been a crucial topic in the medical field. However, manual diagnosis suffers a series of challenges, such as considerable time cost and low diagnosis consistency. In this paper, we proposed TransRAM, which merited both Transformers and Recurrent Models of Visual Attention, as a strong alternative for digital pathological diagnosis of the breast. This model imitated the diagnosis process of pathologists, paying attention to a series of local features on the pathology image and summarizing related information to arrive at the final diagnosis. TransRAM was evaluated on the Bioimaging Challenge 2015 Breast Histology Dataset (BACH) and achieved a classification accuracy of 90.94%, which was better than related publications. In addition, we compared the model effects under different parameter settings and visualized the model decision-making processes.