Project 05 / Computer vision

Pneumonia detection

Making model predictions inspectable.

Overview

A closer look
at the system.

A chest X-ray classification project with two-stage transfer learning. Grad-CAM overlays help inspect the model’s image features alongside its predictions.

Engineering focus: Computer vision
What this demonstrates: Transfer learning, model evaluation, Grad-CAM visualization, and model serving.
Project context: Computer vision & model serving
Technology: PyTorch / EfficientNet-B0 / FastAPI / Streamlit / ONNX

System study / 05Conceptual flow

01
Chest X-rayPyTorch
02
Classify imageEfficientNet-B0
03
Inspect attentionGrad-CAM
04
Serve predictionFastAPI · ONNX

Computer vision

Model results

Predictions, made visible.

Four chest X-ray classification examples with original images, normal or pneumonia labels, model confidence values, and Grad-CAM overlays — 01 / Prediction examples and Grad-CAM overlaysView full size ↗

01 / The problem

An image classification project needs both evaluation and a way to inspect which image regions influence predictions.

02 / Implementation & design

An ImageNet-pretrained EfficientNet-B0 is trained in two stages: a frozen backbone with a new classifier, followed by fine-tuning at a lower learning rate. Grad-CAM overlays support visual inspection, while FastAPI inference and ONNX export explore model serving.

03 / Engineering challenge

Class imbalance and a small original validation set motivated weighted loss and a stratified training/validation split. The repository reports 98.2% recall with 88.0% precision at a 0.5 threshold on 624 test images. These are project evaluation results, not clinical validation.

Results & capabilities

98.2%

Reported recall on project evaluation

90.5% accuracy · 88.0% precision · 624 test images

Metrics from the existing project evaluation; not an independent benchmark.