Related Experiment Video
Updated: Aug 9, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Generating post-hoc explanation from deep neural networks for multi-modal medical image analysis tasks
Weina Jin1, Xiaoxiao Li2, Mostafa Fatehi3
1School of Computing Science, Simon Fraser University, Burnaby, BC, V5A 1S6, Canada.
Methodsx
|February 16, 2023
Summary
Explaining decisions from deep neural network (DNN) models using multi-modal medical images is crucial for clinical use. This study adapts feature attribution methods to explain DNN predictions on diverse medical image data, providing implementation details.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep neural networks (DNNs) are increasingly used for medical image analysis.
- Explaining DNN decisions is essential for clinical adoption as decision support tools.
- Multi-modal medical images offer complementary information for diagnosis.
Purpose of the Study:
- To adapt and implement post-hoc artificial intelligence feature attribution methods for explaining DNN decisions on multi-modal medical images.
- To provide a framework for understanding how DNNs utilize information from different imaging modalities.
Main Methods:
- Utilized gradient-based explanation methods (e.g., Guided BackProp, DeepLift).
- Employed perturbation-based methods (e.g., occlusion, LIME, kernel SHAP).
- Described implementation details for applying these methods to multi-modal image inputs.
Main Results:
- Demonstrated the applicability of established feature attribution techniques to multi-modal medical imaging data.
- Provided insights into how DNN models interpret and integrate information from various image sources.
- Made implementation code publicly available for reproducibility and further research.
Conclusions:
- Feature attribution methods can effectively explain DNN decisions on multi-modal medical images.
- This work facilitates the clinical deployment of AI-driven decision support systems.
- Open-source implementation promotes transparency and advancement in medical AI explainability.
