Understanding Anatomy Classification Through Attentive Response Maps

TL;DR

This paper introduces a visualization method for CNNs using attentive response maps, enabling better understanding of model decisions and alignment with medical landmarks, which aids in transparency and bias detection.

Contribution

The paper proposes a novel visualization technique for CNN internal activations, improving interpretability and alignment with domain-specific landmarks in medical imaging.

Findings

Deep CNNs can be trained to focus on known medical landmarks.

Attentive response maps facilitate understanding of model decisions.

The approach helps in detecting biases in models.

Abstract

One of the main challenges for broad adoption of deep learning based models such as convolutional neural networks (CNN), is the lack of understanding of their decisions. In many applications, a simpler, less capable model that can be easily understood is favorable to a black-box model that has superior performance. In this paper, we present an approach for designing CNNs based on visualization of the internal activations of the model. We visualize the model's response through attentive response maps obtained using a fractional stride convolution technique and compare the results with known imaging landmarks from the medical literature. We show that sufficiently deep and capable models can be successfully trained to use the same medical landmarks a human expert would use. Our approach allows for communicating the model decision process well, but also offers insight towards detecting biases.