Pneumonia classification with a dual-pathway CNN
A 4.2M-parameter CNN trained from scratch. Two branches read the same image at different kernel scales, and each produces its own attention map. Everything below runs in your browser — the image never leaves your device and there is no server.
The architecture was designed to treat disagreement between those branches as its uncertainty signal. Measured against the model's own errors that scores AUROC 0.5334, which is chance, so the model reports predictive entropy instead and shows the disagreement value alongside it as a recorded negative result.
checking for model…
1 · Provide a chest radiograph
Drop an image here
PNG or JPEG, frontal chest radiograph. Processed locally.
No sample images are bundled: the test set is patient radiographs, and redistributing them from this site is a decision for the dataset's terms rather than a convenience. Any frontal chest radiograph works — the RSNA dataset is on Kaggle.
Input as the model sees it
2 · Not read
2 · Prediction
Not a medical device. Research output. This recommendation has no clinical authority and must not inform any diagnosis or treatment decision.
—
The band is conveyed by its label and icon as well as its colour. Thresholds are the 60th and 85th percentile of entropy on the validation split — a property of this model, not a clinical standard.
Pathway disagreement: — — reported, but not used. This architecture was designed to treat disagreement between its two pathways as the uncertainty signal. Measured against the model's own errors it scores AUROC 0.5334 over the full test set on three seeds, which is chance, and combining it with entropy makes entropy worse (0.7675 against 0.7747). Both heads are trained on the same labels with the same loss, so they converge toward each other and rarely disagree at all. The number is shown because the negative result is part of the finding.
3 · What each pathway looked at
Class activation maps, computed without gradients: each pathway head is a single linear layer on globally-pooled features, so the map is a weighted sum of feature maps. If the two pathways attend to the same region, they agree; if not, that disagreement is what the uncertainty score measures.