Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Deep Learning Model Explainability in Brain Tumor Datasets using Post-Heuristic Approaches

Konstantinos Pasvantis, Eftychios Protopapadakis

TL;DR

The paper tackles the explainability gap in brain-tumor detection from MRI by introducing a post-processing refinement of the LIME Image Explainer. It combines transfer-learned CNN classifiers (InceptionV3, ResNet50V2, NasNetLarge) with a model-agnostic explanation framework to generate heatmaps and then refines these explanations using a brain-mask derived from edge-detection methods, retaining segments that predominantly lie inside the brain region. The approach yields improved tumor-segment coverage, notably achieving 50.28% with 3 segments and up to 63.84% with 5 segments, though brain-mask consistency remains a challenge; the best balance was observed with 3 segments. Experimental results show ResNet50v2 as the top classifier among the tested models, and the refinement enhances interpretability in a clinically relevant context, suggesting potential for greater trust and integration into decision-making processes. Future work should focus on more robust brain-mask generation and adaptive thresholds to further stabilize explanations across diverse MRI data.

Abstract

The application of deep learning models in medical diagnosis has showcased considerable efficacy in recent years. Nevertheless, a notable limitation involves the inherent lack of explainability during decision-making processes. This study addresses such a constraint, by enhancing the interpretability robustness. The primary focus is directed towards refining the explanations generated by the LIME Library and LIME image explainer. This is achieved throuhg post-processing mechanisms, based on scenario-specific rules. Multiple experiments have been conducted using publicly accessible datasets related to brain tumor detection. Our proposed post-heuristic approach demonstrates significant advancements, yielding more robust and concrete results, in the context of medical diagnosis.

Enhancing Deep Learning Model Explainability in Brain Tumor Datasets using Post-Heuristic Approaches

TL;DR

The paper tackles the explainability gap in brain-tumor detection from MRI by introducing a post-processing refinement of the LIME Image Explainer. It combines transfer-learned CNN classifiers (InceptionV3, ResNet50V2, NasNetLarge) with a model-agnostic explanation framework to generate heatmaps and then refines these explanations using a brain-mask derived from edge-detection methods, retaining segments that predominantly lie inside the brain region. The approach yields improved tumor-segment coverage, notably achieving 50.28% with 3 segments and up to 63.84% with 5 segments, though brain-mask consistency remains a challenge; the best balance was observed with 3 segments. Experimental results show ResNet50v2 as the top classifier among the tested models, and the refinement enhances interpretability in a clinically relevant context, suggesting potential for greater trust and integration into decision-making processes. Future work should focus on more robust brain-mask generation and adaptive thresholds to further stabilize explanations across diverse MRI data.

Abstract

The application of deep learning models in medical diagnosis has showcased considerable efficacy in recent years. Nevertheless, a notable limitation involves the inherent lack of explainability during decision-making processes. This study addresses such a constraint, by enhancing the interpretability robustness. The primary focus is directed towards refining the explanations generated by the LIME Library and LIME image explainer. This is achieved throuhg post-processing mechanisms, based on scenario-specific rules. Multiple experiments have been conducted using publicly accessible datasets related to brain tumor detection. Our proposed post-heuristic approach demonstrates significant advancements, yielding more robust and concrete results, in the context of medical diagnosis.
Paper Structure (16 sections, 7 equations, 6 figures)

This paper contains 16 sections, 7 equations, 6 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. Research Challenges
  4. Our Contribution
  5. Proposed Methodology
  6. Employed Deep Learning Architectures
  7. Model's Explainability
  8. Post-processing refinement mechanisms
  9. Experimental Setup
  10. Utilized Dataset
  11. Dataset Preprocessing
  12. Performance Metrics
  13. Experimental Results
  14. Advantages and Limitations
  15. Conclusions
  16. ...and 1 more sections

Figures (6)

  • Figure 1: Proposed Methodology
  • Figure 2: Top segments contributing to the prediction, with the produced heatmap from LimeImageExplainer
  • Figure 3: Performance Metrics of each model for each fold. The last collumn corresponds to each model after k-fold validation
  • Figure 4: Average Tumor Segment Coverage for each edge detector regarding the number of segments.
  • Figure 5: Brain Segment Coverage Distributions for each edge detector regarding the number of segments.
  • ...and 1 more figures