FIXUS-APP
A Flask web app that classifies foot and ankle X-rays and overlays Grad-CAM heatmaps to show what the model is looking at.
The problem
X-ray views vary (AP, lateral, oblique), and reading them still leans heavily on the individual radiologist or orthopaedic doctor's experience. A second, explainable opinion can help.
Why it mattered
Automates the first pass of classification and shows, visually, which regions of the image drove the result — rather than handing back a black-box label.
Topics
Built with
Architecture
A Flask app sits in front of an InceptionV3-based classifier trained on foot and ankle X-rays. Grad-CAM generates a heatmap alongside each prediction, so the output is a classification plus a visual explanation of which regions of the image mattered — rather than a label on its own.
Data
Trained and evaluated across the three common X-ray views (AP, lateral, oblique), since the model needs to generalise across viewing angle rather than assume one fixed perspective.
Model
InceptionV3 was used as the base classifier — a pre-trained architecture fine-tuned for this task — paired with Grad-CAM for interpretability, since a classification alone isn’t enough in a diagnostic-support context; showing why matters as much as the label.
Deployment
Runs as a Flask web app, accessible from a browser rather than requiring any local setup.
Lessons
Interpretability (Grad-CAM) turned out to matter as much as raw accuracy for this to be useful to an actual radiologist — a confident wrong answer with no explanation is worse than a visible, checkable one.