Student runs explainable bone-lesion X-ray screener for £5/month

xrY- · reddit · 2026-09-03

A biomedical-to-CS student built an explainable X-ray bone-lesion screener that costs at most £5/month. A DenseNet-121 trained on BTXRD (3,746 radiographs) classifies normal/benign/malignant, using focal loss with inverse-frequency alpha and validation-only threshold tuning to counter the 9.1% malignant class imbalance. It runs on a Cloudflare Worker with a model container that sleeps after 90 seconds and is metered at 5 runtime-hours/month. The author flags open issues: hand-tuned (not learned) OOD gating that rejects non-X-ray uploads, and a single-reviewer gate so user flags never directly poison training labels. Live at osteoneuralnetwork.com.

Original post →

More from coding & agent

coding & agent channel →