Incremental Learning for Disease Classification

Abstract

Current deep-learning models have proven to be effective at the recognition of disease classification in medical images. However, these models are not adapted to the dynamic and ever-changing clinical environment, which sees an influx of new diseases to be classified, and an increase in annotated medical data. To fix this, we implement the use of continual learning systems, which sequentially learn from new samples without forgetting previously learned knowledge. We test models on the MedMNIST dataset, which provides a benchmark for biomedical image classification. We explore rehearsal-based incremental learning and test the use of fixed and growing exemplar memory. An analysis of exemplar memory size on performance is also provided.