Continual knowledge distillation for histopathology

Abstract

With the emergence of computational pathology, many datasets were made public and challenges were published to encourage researches into developing assistant frameworks for pathology tasks. The analysis of histopathological slides, made by pathologists to detect tumorous cells or metastasis in tissue images, is one of such tasks, for which, computer vision had been successfully applied and even outperformed human expert levels. Despite the excellent results in the literature, the majority of approaches are dataset-dependent and lack generalization, making even the best documented models perform poorly when presented with different tissues. In this work, we designed a novel continuous learning method, that leverages the model generalization across datasets using enhanced knowledge distillation. We verified, through deep and extensive experimentation on 19 datasets, an overall improvement of 15,66% in comparison to common literature methods, and superior metrics in relation to models with full dataset availability. Also, our method was the only one to achieve positive forward (FWT) and backward (BWT) knowledge transfer indexes, considerably mitigating the catastrophic forgetting effect.