Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

3D ReX: Causal Explanations in 3D Neuroimaging Classification

Melane Navaratnarajah, Sophie A. Martin, David A. Kelly, Nathan Blake, Hana Chockler

TL;DR

3D ReX introduces a causality-based, post-hoc explainability tool for 3D neuroimaging by leveraging actual causality to produce voxel-level responsibility maps without requiring model weights. The method iteratively occludes supervoxels to identify regions causally responsible for a stroke classifier’s decision, using a partitioning strategy that keeps computation tractable. Empirical results on 3D MRI data show that a $0$-value occlusion yields explanations aligned with lesion locations, while alternative occlusions reveal broader brain regions, offering potentially clinically meaningful and novel insights. The work acknowledges computational and validation limitations and outlines future directions toward multi-modal extensions and standardized evaluation frameworks.

Abstract

Explainability remains a significant problem for AI models in medical imaging, making it challenging for clinicians to trust AI-driven predictions. We introduce 3D ReX, the first causality-based post-hoc explainability tool for 3D models. 3D ReX uses the theory of actual causality to generate responsibility maps which highlight the regions most crucial to the model's decision. We test 3D ReX on a stroke detection model, providing insight into the spatial distribution of features relevant to stroke.

3D ReX: Causal Explanations in 3D Neuroimaging Classification

TL;DR

3D ReX introduces a causality-based, post-hoc explainability tool for 3D neuroimaging by leveraging actual causality to produce voxel-level responsibility maps without requiring model weights. The method iteratively occludes supervoxels to identify regions causally responsible for a stroke classifier’s decision, using a partitioning strategy that keeps computation tractable. Empirical results on 3D MRI data show that a -value occlusion yields explanations aligned with lesion locations, while alternative occlusions reveal broader brain regions, offering potentially clinically meaningful and novel insights. The work acknowledges computational and validation limitations and outlines future directions toward multi-modal extensions and standardized evaluation frameworks.

Abstract

Explainability remains a significant problem for AI models in medical imaging, making it challenging for clinicians to trust AI-driven predictions. We introduce 3D ReX, the first causality-based post-hoc explainability tool for 3D models. 3D ReX uses the theory of actual causality to generate responsibility maps which highlight the regions most crucial to the model's decision. We test 3D ReX on a stroke detection model, providing insight into the spatial distribution of features relevant to stroke.

Paper Structure

This paper contains 7 sections, 2 figures, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Experiments
  5. Discussion
  6. Limitations
  7. Future Work

Figures (2)

  • Figure 1: Visualization of the model's explanation for patient A, predicted as stroke with a confidence of $0.9756$. The purple outline represents the annotated lesion location from the dataset, while the orange outline represents the explanation generated by 3D-ReX. The patient's brain is also overlaid with the responsibility map generated by 3D-ReX, and the color bar depicts the normalized intensity of feature contributions.
  • Figure 2: 3D rendering of patient A's results showing the patient's brain with overlaid annotated lesion location from the dataset (purple) and 3D-ReX's explanation (orange). This visualization provides a spatial perspective of the explanation.