README.md

metadata

title: NeuroSSL Dementia Classifier
emoji: 🧠
colorFrom: indigo
colorTo: purple
sdk: docker
app_port: 7860
pinned: true
license: mit
short_description: Self-Supervised ViT for MRI-based Dementia Classification

NeuroSSL — Dementia Classifier

A deep learning clinical diagnostics app built with a Self-Supervised Vision Transformer (ViT) trained on OASIS & IXI brain MRI datasets.

Features

• Real-time Inference on brain MRI slices with attention rollout visualization
• Platt-calibrated probabilities with MC Dropout uncertainty quantification
• Interactive MRI Viewer with pan, zoom, and multiple scientific colormaps
• Diagnostic Report Generator with PDF/print export
• Model Analytics Dashboard with cross-validation metrics and training curves

Architecture

• Encoder: MultiScaleViT2D (384-dim, 12 blocks, 6 heads)
• Classifier: Second-Order Covariance Pooling
• Calibration: Platt Temperature Scaling (T = 1.6995)
• Uncertainty: Monte Carlo Dropout