Pathology Image Classification for Colorectal Cancer Prediction Using Experience Replay

Pathology image classification helps diagnosis and monitoring as well as treatment of a disease. Thus, it can help the decision-making of a health worker. Still, conventional machine learning models struggle with medical data that often changes, as they typically need retraining on the full dataset to learn new information on a frequent basis. This retraining requires a lot of computations and sometimes loses what it “knew” previously. This is called catastrophic forgetting. To solve this problem, we propose a continual learning framework, using Experience Replay to gradually add new pathology image data, while retaining knowledge of older tasks. Using Experience Replay, the model replays a sample of past data while training so that it does not forget any new information. The proposed framework is validated on the PathMNIST dataset, where it shows better classification accuracy and robustness than the classical retraining. This indicates that continual learning has the potential to be used for pathology image classification, and as a result, it can be scalable and adaptive to the needs of large-scale medical data. Through such advancements diagnostic accuracy will increase, resource use will become more efficient, and patient outcomes will improve in medicine.