HydraGCN

Multi-Modal data analysis framework for medical applications

GCN Multi-Modality Alzeihmer_Prediction COVID-19_Classification Skin_Lesion_Detection

About

In medical research, multi-modal datasets from large-scale population-based studies are an essen:al tool towards better diagnosis and treatment of disease. Multimodal data comprises imaging and non-imaging data and is available in many domains. In technical research, such datasets serve as enablers of Computer-Aided Diagnosis (CADx) with machine learning (ML). In this project, we will be focusing on four such multi-modal dataset HCP dataset and UK Biobank for the application of age and gender predictoon. In this project, we analyze multi-modal datasets with all its challenges such as imbalance, small size, missing values etc with the perspec:ve of graph convolu:onal networks. We will work on a framework developed at CAMP called *‘HydraGCN’* which is designed with a multi-head input and output architecture, which allows for flexible modeling of multimodal input data and which enables multi-target learning through separate output heads and losses.

Task	Classes	Dataset	Modality
Alzeihmer Disease Prediction	CN - Cognitively normaL,MCI - mild cognitive impairment or AD - probable Alzheimer's Disease	TADPOLE	Clinical Features
Covid-19 Disease Classification	Covid Positive or Negative	Covid-19 iCTCF	Clinical Features + Images
Skin Lesion Detection	Melanocytic nevi, Melanoma, Benign keratosis-like lesions, Basal cell carcinoma, Actinic keratoses, Vascular lesions, Dermatofibroma	HAM10K	Clinical Features + Images

Graph Convolutional Networks

We have used an ensemble of 4 GCN's

• GCN - Standard GCN
• MGMC - Multi-Graph Matrix Completion
• GAT - Graph Attention Networks
• DGM - Differentiable Graph Module

Setup

• Create a conda env from the requirements file

conda create -y --name hydragcn python=3.7
conda install -c conda-forge --file requirements.txt
conda activate hydragcn

• Create Dataloader

Update for each dataset the classes in /base/setup/dataset_setup.py

• Update the Configurations

Go to /application/loop_experiments.py

# Choose one dataset to work on

dataset_list = ['TADPOLE']
dataset_list = ['COVIDiCTCF']
dataset_list = ['HAM10K']

# Choose the model ensemble to work on

model_list = ['GCN']
model_list = ['FullGAT']
model_list = ['MGMC']
model_list = ['CDGM']
model_list = ['Hydra']

# Choose the corresponding yaml configuration OR make your own @'../base/configs/<dataset>/<model_you_like>.yaml'

model_yaml_dir = '../base/configs/TADPOLE/MGMC.yaml'
model_yaml_dir = '../base/configs/COVID/MGMC.yaml'
model_yaml_dir = '../base/configs/COVID/Hydra.yaml'
model_yaml_dir = '../base/configs/HAM10K/Hydra_HAM10K.yaml'

Run the Experiment

# Starting point of our code 
python /applications/loop_experiments.py

Problems ? Happy Debugging !