Keywords: Digital Pathology, Continual Learning, Tumor Tissue Classification
TL;DR: We performed the first systematic study to assess the effectiveness of continual learning models in digital pathology applications.
Abstract: Deep-learning models have achieved unprecedented performance in fundamental computational tasks in digital pathology (DP) based analysis, such as image classification, cell detection and tissue segmentation. However, such models are known to suffer from catastrophic forgetting when adapted to unseen data distribution with transfer learning. With an increasing need for deep-learning models to handle ever-changing data distributions, including evolving patient population and new diagnosis assays, it is crucial to introduce methods for alleviating the such model forgetting. To this end, continual learning (CL) models are promising candidates. However, to our best knowledge, there’s no systematic study of CL models in DP-specific applications. Here, we propose various CL scenarios in DP settings, where histopathology image data from different sources/distributions arrive sequentially, the knowledge of which is integrated into a single model without training all the data from scratch. To benchmark the performance of recently proposed continual learning algorithms in the proposed CL scenarios, We augmented a dataset for colorectal cancer H&E classification to simulate shifts of image appearance and evaluated CL methods on this dataset. Furthermore, we leveraged a breast cancer H&E dataset along with the colorectal cancer dataset to assess continual learning methods for learning from multiple tumor types. We revealed promising results of CL in DP applications, potentially paving the way for application of these methods in clinical practice.