REDDEL-Net

Description:

A Renyi Entropy-Driven Deep Ensemble Learning Approach for Classifying Alzheimer's Disease from Structural Brain MR Images

Methodology

• A lightweight, interpretable, hybrid learning framework for Alzheimer's Disease classification using brain MRI scans
• It employs fine-tuned MobileNetV2 with embedded Convolutional Block Attention Module (CBAM) blocks after every convolutional 2d layer
• The trained MobileNetV2+CBAM model is used as a feature extractor; Grad-CAM and LIME are applied to visually highlight the most informative regions in MRI slices
• Renyi-entropy based feature selection optimizes the process by using 7-19% of the extracted feature space by preserving key discriminative information
• Soft-voting ensemble of Logistic Regression, Support Vector Machine with Radial Basis Function, and Random Forest is used for classification
• t-Stochastic Neighborhood Embedding is used to map the distinct class-wise separation; Confidence Intervals (>95%) are illustrated using violin plots
• Achieves 99.38% and 99.89% F1-Score on Kaggle AD and OASIS-1 datasets respectively
• REDDEL-Net requires minimal computational resources and inference times, making it suitable for real-time deployment in clinical environments

Datasets Used

• Kaggle AD
• OASIS-1

End-to-End Workflow Used for Proposed Framework:

Results for Kaggle AD:

• Training & Validation Loss Curve for MobileNetV2+CBAM over Epochs for Kaggle AD

• Feature Selection versus Accuracy for Kaggle AD

• Accuracy curve for 5-fold Stratified Cross Validation for Kaggle AD

• Violin Plot for Distribution of Performance Metrics Across Cross Validation Folds for Kaggle AD

• System Resource Usage Graph During Inference of Test Set on Kaggle AD

• Confusion Matrix for Kaggle AD

• t-SNE of Kaggle AD Test-Set Features by Predicted Labels

XAI Visualizations for OASIS-1 and Kaggle AD

• Grad-CAM for OASIS-1 and Kaggle AD

• LIME for OASIS-1 and Kaggle AD

• SHAP for OASIS-1 and Kaggle AD

Installation

To install the required packages, run:

pip install -r requirements.txt

Program Files:

• Dependencies
• Data Loader
• MobileNetV2+CBAM Architecture
• Training
• Evaluation
• Feature Extraction
• Renyi-Entropy + ML-Ensemble Classifer
• Explainable AI (XAI)

Scripts:

• Training Pipeline
• ML Pipeline
• Inference Pipeline

Implementation Files:

• Kaggle AD Implementation Jupyter Notebook
• OASIS-1 Implementation Jupyter Notebook

Importance of Project:

• Exemplary results with minimal computational resources and inference times
• End-to-end interpretability and outperforms SOTA approaches on benchmarked, standard datasets
• Real-time clinical applicability for better therapeutic outcomes

Paper:

It'd be great if you could cite our paper (under review) if this code has been helpful to you.

Thank you very much!