Case study · 2024

Radiography Image Classification

A multi-class chest X-ray classifier paired with an interpretability dashboard for non-technical review.

Discipline
Machine learning · Computer vision
Technology
TensorFlow · Keras · MobileNetV2 · Streamlit · Grad-CAM
GRAD-CAM / MODEL ATTENTION

I trained a transfer-learning pipeline on approximately 17,000 chest X-ray images, then built a Streamlit interface that exposed both predictions and Grad-CAM attention regions.

≈17K imagesTraining corpus
MobileNetV2Base architecture
Focal LossImbalance strategy
Grad-CAMInterpretability

How the pieces connect

01≈17K X-rays
02MobileNetV2 · Focal Loss
03Prediction + Grad-CAM

What made the problem worth solving

Radiography datasets are difficult for two connected reasons: classes are not distributed evenly, and a model score alone does not show a reviewer why the network made its decision.

Turning the problem into an engineering plan

  1. 01

    Used MobileNetV2 transfer learning to start from strong visual features instead of training an entire convolutional network from scratch.

  2. 02

    Applied Focal Loss so difficult and underrepresented examples had more influence during training.

  3. 03

    Added Grad-CAM overlays in a Streamlit dashboard to surface the image regions contributing to a prediction.

The project became more than a model notebook: it was a reviewable ML system with a clear path from input image to prediction and visual explanation.

Next case studySTEM Forge