Medical Imaging

X-ray Body Part Classifier – Quick Demo using AI

🧩 Problem Statement

In hospitals, thousands of X-ray images are generated daily but many are:

• Unlabeled or misfiled
• Mixed in storage without body-part tags

This makes it hard for doctors and AI systems to automatically route or diagnose these images efficiently.

---

❓ Why Are We Solving This?

Automatically identifying which body part an X-ray image belongs to can:

• Help organize and sort large medical datasets
• Serve as a preprocessing step for disease detection models
• Reduce human effort in data labeling
• Allow smart routing of images to relevant specialists

---

Approach

We are building a lightweight AI model that:

1. Takes any X-ray image as input
2. Predicts the body part shown (e.g., wrist, shoulder, humerus)
3. Displays the prediction in a simple UI

---

Dataset

For this demo, we used a subset of the MURA dataset (Stanford):

• 7 body part classes: elbow, finger, forearm, hand, humerus, shoulder, wrist
• 200 images per class to keep training fast

Structure:

data/ └── train/ ├── elbow/ ├── hand/ ├── shoulder/ └── …

Public datasets:

• MURA Dataset (Stanford)
• VinDr X-ray Dataset (Kaggle)

---

Tech Stack

Component	Tool
Deep Learning Framework	PyTorch
Model	Pretrained ResNet18
Frontend UI	Streamlit
Image Processing	Pillow, TorchVision
Deployment	Local browser via streamlit run

---

Working

1. Training a ResNet18 on your labeled X-ray images
2. Saving the model and class labels
3. Using Streamlit to create a web app:
   • Upload an X-ray image
   • View predicted body part and confidence scores

---

Output

• xray_bodypart_model.pth – the trained model
• Live web app to classify uploaded X-rays by body part

---