Proptosis Medical Screening Computer Vision Project

This repository hosts the code, datasets, and documentation for the independent study titled "Leveraging Self-Supervised Learning for Early Medical Screening: A Pilot Study in Proptosis Detection." The project investigates the use of self-supervised learning techniques to detect subtle facial indicators of proptosis, a potential early sign of systemic diseases such as Thyroid Eye Disease and tumors. The study is organized into two main phases: a proof-of-concept using synthetic or toy datasets, followed by the application of these methods to clinical proptosis screening data (with all necessary ethical approvals and safeguards in place).

Project Overview

• Objective: To assess the effectiveness of self-supervised learning in low-data scenarios and to develop a non-invasive early screening tool for proptosis.
• Phases:
  1. Proof-of-Concept: Validate the self-supervised approach on a controlled toy dataset (e.g., animal images) to establish baseline performance and robust evaluation metrics.
  2. Clinical Application: Fine-tune the model on anonymized clinical images for proptosis detection and compare its performance against supervised baselines.
• Team:
  • Officially Pursuing Independent Study:
    • Aryan Sajith (3 credits)
  • Additional Contributors
    • Dhriti Madireddy
    • Aaditya Saini
    • Grace Zhou
• Advisor: Professor Erik Learned-Miller
• Semester: Spring 2025

For complete project specifications, please refer to the Project Proposal

Repository Structure

• code/ — Contains code for model implementations, training scripts, and so on.
• docs/ - Contains important documents like the project proposal.
• README.md — The primary file that outlines the details of the independent study.

Setup and Usage

TBD - Will be setup as project progresses.

Phases

• Phase 1: Literature Review — Review self-supervised learning methods and proptosis detection literature.
• Phase 2: Toy Problem Experimentation — Train models on synthetic datasets to validate the methodology.
• Phase 3: Clinical Application — Fine-tune the approach on anonymized proptosis data.
• Phase 4: Synthesis & Summary — Compile insights and report findings.