Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Explainable concept mappings of MRI: Revealing the mechanisms underlying deep learning-based brain disease classification

Christian Tinauer, Anna Damulina, Maximilian Sackl, Martin Soellradl, Reduan Achtibat, Maximilian Dreyer, Frederik Pahde, Sebastian Lapuschkin, Reinhold Schmidt, Stefan Ropele, Wojciech Samek, Christian Langkammer

TL;DR

This study addresses the interpretability of CNNs for MRI-based Alzheimer's disease classification by using CRP to map learned concepts to brain regions via quantitative $R_2^*$ maps. A Graz^+ relevance-guided CNN is trained on $R_2^*$ data, with a relevance-regularization term that concentrates attention within brain tissue, and concept analysis identifies eight last-layer channels whose relevance is ranked and interpreted. Results reveal that highly relevant concepts concentrate in and around the basal ganglia, with ROI analysis showing corroborating $R_2^*$ differences in subcortical structures, supporting the biological plausibility of learned features. The approach demonstrates that concept mappings can validate models, reveal underlying mechanisms, and potentially improve reliability in clinically relevant MRI classification tasks.

Abstract

Motivation. While recent studies show high accuracy in the classification of Alzheimer's disease using deep neural networks, the underlying learned concepts have not been investigated. Goals. To systematically identify changes in brain regions through concepts learned by the deep neural network for model validation. Approach. Using quantitative R2* maps we separated Alzheimer's patients (n=117) from normal controls (n=219) by using a convolutional neural network and systematically investigated the learned concepts using Concept Relevance Propagation and compared these results to a conventional region of interest-based analysis. Results. In line with established histological findings and the region of interest-based analyses, highly relevant concepts were primarily found in and adjacent to the basal ganglia. Impact. The identification of concepts learned by deep neural networks for disease classification enables validation of the models and could potentially improve reliability.

Explainable concept mappings of MRI: Revealing the mechanisms underlying deep learning-based brain disease classification

TL;DR

This study addresses the interpretability of CNNs for MRI-based Alzheimer's disease classification by using CRP to map learned concepts to brain regions via quantitative maps. A Graz^+ relevance-guided CNN is trained on data, with a relevance-regularization term that concentrates attention within brain tissue, and concept analysis identifies eight last-layer channels whose relevance is ranked and interpreted. Results reveal that highly relevant concepts concentrate in and around the basal ganglia, with ROI analysis showing corroborating differences in subcortical structures, supporting the biological plausibility of learned features. The approach demonstrates that concept mappings can validate models, reveal underlying mechanisms, and potentially improve reliability in clinically relevant MRI classification tasks.

Abstract

Motivation. While recent studies show high accuracy in the classification of Alzheimer's disease using deep neural networks, the underlying learned concepts have not been investigated. Goals. To systematically identify changes in brain regions through concepts learned by the deep neural network for model validation. Approach. Using quantitative R2* maps we separated Alzheimer's patients (n=117) from normal controls (n=219) by using a convolutional neural network and systematically investigated the learned concepts using Concept Relevance Propagation and compared these results to a conventional region of interest-based analysis. Results. In line with established histological findings and the region of interest-based analyses, highly relevant concepts were primarily found in and adjacent to the basal ganglia. Impact. The identification of concepts learned by deep neural networks for disease classification enables validation of the models and could potentially improve reliability.
Paper Structure (17 sections, 4 equations, 5 figures, 2 tables)

This paper contains 17 sections, 4 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset
  4. Preprocessing
  5. Standard classification network
  6. Relevance-guided classification network
  7. Training
  8. Model selection
  9. Bootstrapping analysis
  10. Concepts identification
  11. Relevance-weighted concept map representation
  12. R2* and relevance region of interest analysis
  13. Results
  14. Discussion
  15. Conclusion
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Top row shows schematic overview of conventional back-propagation heat mapping propagating relevance scores backwards through the network creating a single heat map. Line thickness symbolizes the (relative) amount of relevance (in red) flowing through the connections. In contrast, shown in the bottom row, by conditioning on a concept encoded by a hidden-layer channel (highlighted in blue and green), Concept Relevance Propagation (CRP) and RelMax allow to compute concept-conditional explanations and provide semantic meaning for latent model structures, disentangling the learned and identified image features. NC normal control; AD Alzheimer's disease; R relevance.
  • Figure 2: Row (1) shows the difference of conventional mean global heat maps created for AD and NC using LRP-$z^+$-rule for five slices (columns), overlaid on the MNI152 standard-brain 1mm template. In comparison, rows (2) to (5) show the differences of the four most important concepts on the same slices, ranked by their relative importance (percentage next to name) for the classification results. Negative differences are presented as blue-lightblue and indicate regions in concepts with more attribution in AD compared to NC, whereas positive differences, shown as red-yellow, highlight regions in concepts with more attribution in NC compared to AD. If the attribution in NC and AD is similar (e.g. concept is nearly equally used for both classes) the regions disappear. Images are shown in standard-radiological view, thus left and right is flipped and white lines outline the basal ganglia.
  • Figure 3: Rows (1) to (4) show the 4 most important concepts, ranked by their relative importance (percentage next to name) for the classification results. Columns (1) to (5) show the 5 top ranked individual test images for each concept on the same transversal slice. Positive relevance, shown as red-yellow, highlight regions in concepts with higher attributions. White lines in slices outline the basal ganglia. Images are shown in standard-radiological view, causing the left and right side of the brain to be flipped.
  • Figure 4: Row (1) shows the conventional mean global heat maps created for NC using LRP-$z^+$-rule for 5 slices (columns), overlaid on the MNI152 standard-brain 1mm template. In comparison, rows (2) to (5) show the 4 most important concepts on the same slices, ranked by their relative importance (percentage next to name) for the classification results. Positive relevance, shown as red-yellow, highlight regions in concepts with higher attributions. White lines in slices outline the basal ganglia. Images are shown in standard-radiological view, causing the left and right side of the brain to be flipped.
  • Figure 5: Row (1) shows the conventional mean global heat maps created for AD using LRP-$z^+$-rule for 5 slices (columns), overlaid on the MNI152 standard-brain 1mm template. In comparison, rows (2) to (5) show the 4 most important concepts on the same slices, ranked by their relative importance (percentage next to name) for the classification results. Positive relevance, shown as red-yellow, highlight regions in concepts with higher attributions. White lines in slices outline the basal ganglia. Images are shown in standard-radiological view, causing the left and right side of the brain to be flipped.