Patient-Level Alzheimer’s Disease Classification Using OCT

Inspiration

This project was inspired by a recent and unique method for AD classification using retinal OCT, which is a non-invasive imaging method using reflected light to create pictures of the back of eyes.

What it does

Takes in datasets that feature Alzheimer’s Patients and control groups’ Light and Dark retinal OCT scans and classifies the data into two groups(Positive and Negative group). For short, it outputs a single diagnosis probability for each patient.

How we built it

We designed a system to perform classification at patient level using functionally paired Light and Dark OCT scans. For every light-dark pair, we computed a reflectivity difference map(delta) to explicitly model the retina’s response to light changes, which may reveal Alzheimer’s patterns. All light-dark and delta images were processed through one shared Swin Transformer - tiny encoder to enforce consistent representation learning and structural comparison through modalities. We then fused each light-dark pair using an MLP that learns both shared features and their functional difference. The pair embeddings were aggregated into a single patient representation using attention-based MIL pooling and the classifier outputs one Alzheimer’s probability per data with regularization to avoid overfitting.

Challenges we ran into

We faced several challenges while building the classifier. The dataset used a binary .hdr/.img format without readable metadata, which caused shape inference failures and required custom fallback logic for robust image reshaping. We also encountered multiple time zone issues since all the teammates are from a different timezone. Also, the dataset was extremely sparse, containing only 28 Control and non-Control group patients. Limited GPU resources and multiple runtime errors caused by environment conflict and image size mismatch was a big challenge too. Our team used Colab Pro promotions for students to handle resource problems and resolved errors by debugging.

Accomplishments that we're proud of

Designed a new model structure that comprehends light, dark, delta(light-dark) OCT image inputs, explicitly modeling inter-modality differences to enhance disease relevant feature learning and stability across imaging conditions. This approach can improve robustness and interpretability in downstream OCT analysis and applications.

What we learned

“The data is King”. Although we tried numerous hyperparameter tunings and applied fancy algorithms, the key to a successful model was abundant data. The metrics remained the same or even worse although we tried to use various methods to improve. For example, before applying class weights to minimize the imbalance in the data between AD and non-AD, AUROC was 0.875. However, after applying it, AUROC dropped to 0.815. This taught us that we should obtain enough data before diving into fancy algorithms.

What was the limitation?

The primary limitation was the limited availability of OCT datasets labeled with AD. Due to this limitation, the model evaluation was extremely constrained in terms of statistical power and generalizability.

What's next for Patient-Level Alzheimer’s Disease Classification Using OCT

To validate model performance more precisely, we plan to expand the OCT dataset with more AD cases. While publicly available AD-labeled OCT datasets remain limited, several open OCT datasets containing normal controls are available on Hugging Face. These datasets can be leveraged to construct a stronger baseline distribution and to generate augmented OCT images that improve model robustness.

Gen AI

We used generative AIs to look for research related to our project and coding assistance to accelerate our development.

Built With

• numpy
• pandas
• python
• pytorch
• scikit-learn
• timm