spec · pneumoscan-ai · shipped · 2024
PneumoScan AI
Explainable pneumonia detection from chest X-rays
- Role
- Sole engineer · graduation project
- Domain
- Computer Vision / Medical
- Year
- 2024
- Status
- shipped
stack →TensorFlowKerasGrad-CAMStreamlitOpenCV
A CNN classifier for chest radiographs that reports 98.42% test accuracy and overlays a Grad-CAM heatmap so a clinician can see which region drove the prediction.
01 · Problem
A raw classifier score is not usable in a clinical setting — a reviewer needs to know whether the model looked at lung tissue or at a scanner artifact.
02 · Approach
- 01Transfer-learned convolutional backbone with augmentation tuned for radiograph contrast and orientation.
- 02Grad-CAM layer to produce per-image saliency overlays alongside the class probability.
- 03Held-out test split kept strictly separate from tuning; class balance checked at every stage.
- 04Medical-grade Streamlit interface: upload, prediction, confidence, and heatmap in one view.
03 · Outcome
- 98.42% accuracy on the held-out test set.
- Saliency overlays make each prediction auditable rather than a black box.