Overcoming Interpretability and Accuracy Trade-off in Medical Imaging

Ivaxi Sheth, Samira Ebrahimi Kahou

Keywords: Interpretability, concept based explanations

Abstract: Neural networks are considered black boxes. Deploying them into the healthcare domain poses a challenge in understanding model behavior beyond final prediction. There have been recent attempts to establish the trustworthiness of a model. Concept learning models provide insight into the model by introducing a bottleneck layer before the final prediction. They encourage interpretable insights into deep learning models by conditioning final predictions on intermediate predictions of explainable high-level concepts. However, using concept-based models causes a drop in performance which poses an accuracy vs explainability trade-off. To overcome this challenge we propose Coop-CBM, a novel concept learning model. We validate the performance of Coop-CBM on diverse dermatology and histopathology images.

TL;DR: We propose concept based model coop-CBM that is both interpretable and accurate.