Related Experiment Videos
Visual Explanations From Deep 3D Convolutional Neural Networks for Alzheimer's Disease Classification
Chengliang Yang1, Anand Rangarajan1, Sanjay Ranka1
1Dept. of Computer & Information Science & Engineering, University of Florida, Gainesville, FL 32611, USA, ximen14@ufl.edu, anand@cise.ufl.edu, ranka@cise.ufl.edu.
AMIA ... Annual Symposium Proceedings. AMIA Symposium
|March 1, 2019
Summary
We developed three methods to visualize how 3D convolutional neural networks (3D-CNNs) identify Alzheimer's disease in brain scans. These techniques help understand the network
Area of Science:
- Neuroimaging
- Artificial Intelligence
- Medical Diagnostics
Background:
- Alzheimer's disease (AD) diagnosis relies on identifying subtle brain changes.
- 3D convolutional neural networks (3D-CNNs) show promise for AD classification from neuroimaging data.
- Understanding the decision-making process of these complex models is crucial for clinical trust and validation.
Purpose of the Study:
- To develop and evaluate efficient methods for generating visual explanations from 3D-CNNs applied to Alzheimer's disease classification.
- To assess the ability of different visualization techniques to identify key brain regions implicated in Alzheimer's disease.
- To compare the strengths and limitations of various explanation generation approaches for 3D-CNNs in neuroimaging.
Main Methods:
- Developed three distinct approaches for visual explanation generation from 3D-CNNs.
- One method utilized sensitivity analysis on hierarchical 3D image segmentation.
- Two methods focused on visualizing network activations on spatial maps of the brain.
Main Results:
- All three developed methods successfully identified important brain regions for Alzheimer's disease diagnosis, confirmed by visual inspection and a quantitative localization benchmark.
- Sensitivity analysis struggled with the diffuse nature of the cerebral cortex.
- Activation visualization methods were limited by the resolution of convolutional layers.
- The complementary nature of the methods provided diverse perspectives on 3D-CNN behavior.
Conclusions:
- Multiple visualization techniques can effectively highlight critical brain areas for Alzheimer's disease detection using 3D-CNNs.
- Each method possesses unique strengths and weaknesses, suggesting a need for complementary approaches for comprehensive model understanding.
- These visualization tools enhance interpretability and trust in AI-driven Alzheimer's disease classification systems.