Breast Cancer Detection Model Training Strategy Based on Continual Learning

Breast cancer is the most diagnosed cancer but also the most curable cancer if detected early. Convolutional neural network (CNN) is widely adopted in breast cancer diagnosis for its efficiency and accuracy. However, due to its structure and features, catastrophic forgetting and inaccuracy are inevitable in the diagnostic process when the original CNN is used to deal with future tasks based on the previous learning experience. To make things worse, the available dataset of breast tumour samples is very limited, bringing challenges to CNN classification models’ accuracy when encountering new samples. In our research, we included three types of strategies for processing datasets and embedded continual learning method EWC in several popular CNN structures. By comparing various parameters in different strategies and structures, we tested the most accurate learning model we introduced. Surprisingly, our approach turns out to be more accurate and adaptable to the classification of microscopic new images of breast tumour tissue into benign and malignant ones. The experimental outcome shed light on a new research direction in improving state-of-the-art CNN meant for diagnostic function.