Expert identification of visual primitives used by CNNs during mammogram classification

TL;DR

This paper introduces an expert-in-the-loop method to interpret CNN internal units in mammogram classification, revealing their detection of meaningful medical visual patterns and supporting explainability of model decisions.

Contribution

It presents a novel expert-guided approach to label CNN units, linking internal representations to medical phenomena in mammogram analysis.

Findings

CNN units detect medically relevant patterns like masses and calcifications.

Expert labeling correlates internal units with specific medical features.

Models can generate explanations supporting classification decisions.

Abstract

This work interprets the internal representations of deep neural networks trained for classification of diseased tissue in 2D mammograms. We propose an expert-in-the-loop interpretation method to label the behavior of internal units in convolutional neural networks (CNNs). Expert radiologists identify that the visual patterns detected by the units are correlated with meaningful medical phenomena such as mass tissue and calcificated vessels. We demonstrate that several trained CNN models are able to produce explanatory descriptions to support the final classification decisions. We view this as an important first step toward interpreting the internal representations of medical classification CNNs and explaining their predictions.