Explaining 3D Computed Tomography Classifiers with Counterfactuals

TL;DR

This paper introduces a memory-efficient method for generating counterfactual explanations for 3D CT scan classifiers, enhancing interpretability in medical imaging by extending 2D techniques to volumetric data.

Contribution

It adapts the Latent Shift counterfactual method to 3D CT scans using a slice-based autoencoder and gradient blocking, addressing challenges of high memory use and limited data.

Findings

Effective counterfactuals for 3D CT classifiers

Memory-efficient approach suitable for high-resolution images

Applicable to clinical phenotype prediction and lung segmentation

Abstract

Counterfactual explanations enhance the interpretability of deep learning models in medical imaging, yet adapting them to 3D CT scans poses challenges due to volumetric complexity and resource demands. We extend the Latent Shift counterfactual generation method from 2D applications to explain 3D computed tomography (CT) scans classifiers. We address the challenges associated with 3D classifiers, such as limited training samples and high memory demands, by implementing a slice-based autoencoder and gradient blocking except for specific chunks of slices. This method leverages a 2D encoder trained on CT slices, which are subsequently combined to maintain 3D context. We demonstrate this technique on two models for clinical phenotype prediction and lung segmentation. Our approach is both memory-efficient and effective for generating interpretable counterfactuals in high-resolution 3D medical imaging.