MEDscan

Inspiration

We created MEDscan to explore how artificial intelligence can support medical imaging research. We were inspired by the possibility of using deep learning to identify patterns in chest X-rays and make AI-powered image analysis more accessible through a simple, interactive web application.

What We Built

MEDscan is a web-based research prototype that classifies chest X-rays into two categories: Normal and Pneumonia. Users can upload an X-ray and view the model's predicted classification and probability scores. Our goal was to combine machine learning with a user-friendly interface while emphasizing transparency and responsible use of AI in healthcare.

How We Built It

We used Python and PyTorch to explore image classification with a ResNet18-based deep learning model and the PneumoniaMNIST dataset. We developed a consistent grayscale image preprocessing pipeline, incorporated temperature scaling for confidence calibration, and used FastAPI to connect the model to our web interface. We also designed the interface to make model outputs and limitations easier to understand.

Challenges We Faced

One of our biggest challenges was ensuring that image preprocessing remained consistent between training and prediction. We also explored how to make confidence scores more meaningful through calibration. Connecting the machine learning model to the web interface introduced additional challenges, including handling image uploads, returning prediction results, and distinguishing real model outputs from simulated examples. These challenges helped us understand that building an AI application requires careful integration, testing, and honest communication about its limitations.

What We Learned

Through MEDscan, we gained experience with deep learning, image classification, API integration, and web development. We also learned that a model's confidence does not automatically mean its prediction is correct. Calibration, validation, and transparency are essential when working on healthcare-related AI projects.

What's Next for MEDscan

We hope to improve MEDscan by conducting more extensive model evaluations, testing on diverse datasets, and refining how the application communicates uncertainty. MEDscan is a research and demonstration prototype, not a substitute for professional medical evaluation or diagnosis.

Built With

• claude
• copilot
• fastapi
• python