Simple Alzheimer's Disease CNN Model - DL4H Final Project

[jsmutum2]

Contributor: Jay Mutum (jsmutum2@illinois.edu)

NetID: jsmutum2

Contribution Type: Model

Paper: Bruningk et al., "Back to the Basics with Inclusion of Clinical
Domain Knowledge — A Simple, Scalable and Effective Model of Alzheimer's
Disease Classification".
https://proceedings.mlr.press/v149/bruningk21a.html

Description

Adds SimpleADCNN, a model reproduction of the simple Alzheimer's Disease CNN
described in the paper. Implements Conv3d -> BatchNorm3d -> ReLU ->
Dropout blocks as in the paper. This is followed by two dense layers.
Checks input validity with tests, as well as construction,
forward pass, etc. For ablation, tests varying configurations
and modifications to the models in the paper on synthetic data.
Because the ADNI dataset used in the paper was not available
(researchers only) synthetic data was used to test that the model
constructed and ran correctly.

The model in the paper is simple, less space intensive, and more
easily interpretable by healthcare professionals.

Files to Review:

• 'pyhealth/models/simple_ad_cnn.py' - Main model implementation
• 'pyhealth/examples/simple_ad_cnn_classification.py' - Example usage
• 'tests/core/test_ad_cnn.py' - Test cases
• 'docs/api/models.rst'
• 'docs/api/models/pyhealth.models.SimpleADCNN.rst'
• 'pyhealth/models/init.py'