Alzheimer detection

Inspiration

Early diagnosis of Alzheimer’s disease remains a major clinical challenge due to subtle and progressive brain changes. I was inspired to explore whether unsupervised AI models could detect these changes by learning what a healthy brain looks like and identifying deviations without relying on labels.

What it does

The project detects Alzheimer-related abnormalities in brain images by learning the distribution of healthy anatomy. It generates anomaly maps and energy scores that highlight regions and samples that deviate from normal brain structure, enabling stage-aware analysis of disease progression.

How I built it

I built a multiscale deep learning model using a ResNet-based feature extractor combined with flow-based density estimation. Features at multiple spatial scales are modeled using normalizing flows, and anomaly scores are computed using quantile-based aggregation to emphasize localized pathological changes.

Challenges I ran into

One of the main challenges was capturing subtle and localized Alzheimer’s patterns without amplifying noise or artifacts. Another challenge was selecting an appropriate anomaly scoring strategy that remained stable while still being sensitive to early-stage disease.

Accomplishments that I was proud of

 - Designing a fully unsupervised Alzheimer detection framework 
 - Achieving stage-dependent anomaly trends without label supervision 
 - Producing interpretable anomaly maps aligned with known neurodegenerative patterns 
 - Successfully validating model behavior using distribution analysis and Q–Q plots

What I learned

I learned that quantile-based anomaly scoring is more effective than mean or max aggregation for detecting localized brain abnormalities. I also gained insight into how flow-based models behave in medical imaging and how statistical tools can validate model assumptions.