Alzheimer's Image Classification Demo

Browser-based TensorFlow.js prototype for classifying brain scan imagery into four dementia-related labels: MildDemented, ModerateDemented, NonDemented, and VeryMildDemented.

This project is a compact example of taking a trained image model, exporting it to TensorFlow.js, and wrapping it in a simple client-side interface that can run without a backend inference server.

What It Shows

• TensorFlow.js graph model served directly from static files
• Four-class image classification UI with local image upload
• Model weights split into browser-loadable shard files
• Lightweight HTML/JavaScript inference flow
• Early ML experimentation around healthcare image classification

Model Summary

Item	Detail
Runtime	TensorFlow.js in the browser
Model format	graph-model
Input	RGB image tensor, dynamic height and width, 3 channels
Output	4 class scores
Classes	MildDemented, ModerateDemented, NonDemented, VeryMildDemented
Weight shards	4 local .bin files

The original project notes reported an average detection result around 90 percent in project testing, with class-level precision notes:

Class	Reported precision
NonDemented	65%
MildDemented	94.40%
VeryMildDemented	97.93%
ModerateDemented	100.00%

Run Locally

The model must be served through HTTP so TensorFlow.js can load model.json and the weight shards.

python -m http.server 8000

Then open:

http://localhost:8000

Use the bundled sample image or upload a similar top-down brain image, then click predict.

Repository Layout

index.html              Browser UI and inference code
model.json              TensorFlow.js graph model manifest
classes.json            Class-to-index mapping
group1-shard*.bin       Model weight shards
image.jpg               Sample input image

Verification

Current automated checks performed:

• Parsed model.json and classes.json
• Confirmed all 4 weight shards referenced by model.json exist
• Served index.html, model.json, classes.json, and all weight shards over local HTTP

Limitations

This is a research and portfolio prototype, not a clinical system or medical device. The repository does not include the original training dataset, a reproducible training pipeline, or independent validation artifacts. Any real healthcare use would require dataset documentation, privacy review, external validation, bias evaluation, and clinician oversight.

Next Improvements

• Add a model card with dataset source, training split, preprocessing, and evaluation method
• Add confusion matrix, recall, F1, and per-class support counts
• Normalize input preprocessing so uploaded images match training conditions
• Add confidence bars instead of returning only the top class
• Add a small test harness that verifies TensorFlow.js inference against the sample image