Inspiration
Alzheimer’s is one of the most devastating neurodegenerative diseases, and early detection is still incredibly challenging. While studying AI for healthcare, I realized that MRI scans contain hidden spatial patterns that even trained clinicians may miss. This inspired me to build a deep-learning system that could support early diagnosis, reduce workload, and potentially help millions of patients receive care sooner.
What it does
The project uses MobileNetV2-based transfer learning to classify preprocessed brain MRI images into Alzheimer’s stages. It includes:
Automated data preprocessing (grayscale → RGB → normalization)
Transfer learning pipeline with MobileNetV2
Data augmentation for robustness
Grad-CAM heatmaps for interpretability
Full training/validation/metrics workflow
The model outputs:
Predicted Alzheimer’s stage
Confidence scores
Visual explanation of what part of the MRI influenced the decision
How we built it
The system was developed using:
Python + TensorFlow/Keras for deep learning
MobileNetV2 pretrained on ImageNet
MRI dataset provided through Hack4Health (converted from parquet)
NumPy / Pandas for data wrangling
Matplotlib for visualization
Grad-CAM for explainability
Pipeline summary:
Loaded MRI slices from parquet files
Converted grayscale MRI → 3-channel RGB
Resized all images to 224 × 224 224×224
Split dataset using stratified sampling
Trained MobileNetV2 with frozen base → fine-tuning
Evaluated using accuracy, F1-score, ROC
Generated Grad-CAM heatmaps for insights
Challenges we ran into
Converting parquet-stored MRI data into usable pixel arrays
Handling grayscale MRI images that needed 3-channel replication
Efficient training on limited GPU resources
Balancing classes to avoid bias
Debugging shape mismatch errors during preprocessing
Ensuring Grad-CAM worked properly on MobileNetV2
Each challenge forced deeper understanding of TensorFlow layers, data pipelines, and model interpretability tools
Accomplishments that we're proud of
Successfully built a high-accuracy MobileNetV2 Alzheimer’s classifier
Integrated Grad-CAM, making the model explainable
Achieved strong performance improvement using augmentation
Created a clean, modular, open-source project others can reuse
Developed a fully reproducible pipeline — from data loading → training → visualization
What we learned
How transfer learning dramatically boosts medical ML performance
Importance of interpretability in healthcare AI
How Grad-CAM works internally using class-specific gradients
How model performance can shift based on augmentation and class imbalance
Best practices for training CNNs on medical images
How to document and structure a real machine learning project
What's next for Alzheimer-Prediction-by-Deep-Learning
Future improvements may include:
Training on larger multimodal datasets (MRI + clinical data)
Using more advanced architectures like EfficientNetV2 or ViTs
Improving segmentation of brain regions before classification
Building a lightweight API for inference
Deploying as a web-based tool for radiologist assistance
Adding ensemble learning for higher diagnostic confidence
Ultimately, the goal is to move this from a hackathon prototype to a trustworthy, clinically useful AI assistant for Alzheimer’s detection.
Log in or sign up for Devpost to join the conversation.