Alzheimer's Stage Detection from Brain MRI (ResNet18)
Fine-tuned ResNet18 for 4-class Alzheimer's stage classification from axial brain MRI slices. Developed as part of SDP_CISHA_05 (7th semester Student Development Project).
Classes
- NonDemented
- VeryMildDemented
- MildDemented
- ModerateDemented
Training Details
- Backbone: ResNet18, ImageNet-pretrained, fine-tuned in 2 phases (frozen head warmup, then unfrozen layer4 + fc)
- Dataset: Kaggle Alzheimer's Dataset (4-class, ~6400 images)
- Class imbalance handling: Offline augmentation (6x) applied to the minority ModerateDemented class + weighted CrossEntropyLoss
- Best validation macro-F1: 0.9775
- Test set accuracy: 0.9927
- Test set macro-F1: 0.9947
⚠️ Important Limitations
- Dataset leakage: This Kaggle dataset is documented in academic literature to contain near-duplicate/augmented images shared across its train and test splits. Reported test metrics should be interpreted as a dataset-specific upper bound, not an estimate of real-world generalization to independent clinical scans.
- Not for clinical use: This model is an academic/educational project and has NOT been validated on independent clinical data, reviewed by medical professionals, or approved for diagnostic use. It must not be used to inform real medical decisions.
- Explainability: Grad-CAM visualizations indicate the model attends to anatomically relevant regions (ventricles, periventricular cortex) rather than scan artifacts, but this does not substitute for clinical validation.
Intended Use
Educational and research demonstration of transfer learning + explainable AI (Grad-CAM) applied to medical image classification.