All projects

MEDISCAN

One network that reads both bone X‑rays and chest X‑rays: it flags fractures and names the chest condition from a single upload.

Client
Clark University capstone
Duration
3 months
Stack
Python, PyTorch, OpenCV, scikit‑learn, Streamlit
MEDISCAN architecture: Streamlit upload, CLAHE preprocessing, DenseNet with fracture and chest condition heads, diagnostic reportMEDISCAN architecture: Streamlit upload, CLAHE preprocessing, DenseNet with fracture and chest condition heads, diagnostic report
System architecture.

MEDISCAN is a dual task diagnosis aid. A DenseNet written from scratch shares one feature extractor between two heads: a fracture detector for bone X‑rays and a 15 class chest condition classifier for chest X‑rays.

On the held out test set the fracture head reached 98.1% accuracy (F1 0.961) and the chest head 70.5% accuracy. A Streamlit app wraps it so anyone can upload a scan and read the result with its confidence.

How it works

  1. Clean every scan the same way

    Grayscale load, CLAHE contrast enhancement, an aspect preserving resize padded to 224 by 224, and pixel values scaled to 0 to 1. Fracture images are augmented five ways to balance the classes.

  2. Train one network on two tasks

    Each batch is half fracture images and half chest images, with "No Finding" chest cases held to a quarter of the chest half so the model cannot coast on the most common label.

  3. Show the answer

    The Streamlit app runs one forward pass, then reports either the fracture verdict or the top three chest conditions with percentages, on CPU or Apple Silicon.

Sequence diagram: user uploads an X‑ray, the image processor and DenseNet model return logits, the decoder shows diagnosis with confidenceSequence diagram: user uploads an X‑ray, the image processor and DenseNet model return logits, the decoder shows diagnosis with confidence
How a request flows through the system.

View the code on GitHub