Related Experiment Video
Updated: May 13, 2025
07:15
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
6.6K
A comparative study of explainability methods for whole slide classification of lymph node metastases using vision
Jens Rahnfeld1,2, Mehdi Naouar1,2, Gabriel Kalweit1,2
1Collaborative Research Institute Intelligent Oncology (CRIION), Freiburg, Germany.
PLOS Digital Health
|April 15, 2025
Summary
Explainable AI methods like ViT-Shapley improve trust in deep learning for pathology. This study found ViT-Shapley offers reliable heatmaps for histopathological images, aiding clinical adoption.
Area of Science:
- Artificial Intelligence
- Digital Pathology
- Medical Imaging
Background:
- Deep learning enhances medical image analysis, particularly in pathology with whole slide imaging.
- Lack of transparency in deep learning models hinders clinical adoption in pathology.
- Explainability methods are crucial for understanding AI predictions in histopathology.
Purpose of the Study:
- Evaluate explainability methods for Vision Transformers in histopathological image classification.
- Compare the effectiveness of Attention Rollout, Integrated Gradients, RISE, and ViT-Shapley.
- Assess heatmap interpretability and reliability for clinical applications.
Main Methods:
- Trained a Vision Transformer on the CAMELYON16 dataset (399 whole slide images).
- Conducted comparative analysis of Attention Rollout, Integrated Gradients, RISE, and ViT-Shapley.
- Evaluated explanation quality using heatmap interpretability and performance metrics.
Main Results:
- Attention Rollout and Integrated Gradients produced artifact-prone heatmaps.
- RISE and ViT-Shapley generated more reliable and interpretable heatmaps.
- ViT-Shapley showed faster runtime and superior performance in insertion/deletion metrics.
Conclusions:
- ViT-Shapley provides reliable and interpretable heatmaps for Vision Transformer predictions on histopathological images.
- Integrating ViT-Shapley heatmaps can enhance trust and scalability in clinical pathology workflows.
- ViT-Shapley facilitates the adoption of explainable AI in digital pathology.
