ARslan-Ahamd/evidential-semi-supervised-lesion-segmentation

Calibrated Uncertainty over Thresholds

Semi-supervised segmentation decides which unlabelled pixels the student learns from, and almost every method decides with a confidence threshold. This page shows why that is the wrong question.

Set the logits to -60 / -60 and then +30 / +30: P(lesion) is 0.500 both times, while vacuity swings from 1.000 to 0.032 and dissonance from 0.000 to 0.968. A threshold discards both pixels; the second was a real boundary.

Reported honestly: a three-seed study retracted this project accuracy claim -- the boundary-F1 gain it was built around is 0.53x the run-to-run noise -- and the one robust statistical result is negative, about its own calibration. What survived is measurement: 32x fewer parameters, 5.87x lower latency, at no measurable accuracy cost.

• —Code: https://github.com/arslan-ahm/evidential-semi-supervised-lesion-segmentation
• —Full results: https://evissl-arslan.surge.sh
• —All seven projects: https://seven-ai-projects-arslan.surge.sh

Runs entirely in your browser through Pyodide — no server, nothing uploaded, nothing leaves your machine. The first load fetches roughly 12 MB and takes 10–40 seconds depending on the demo; after that it is cached.

Research artefact on synthetic data. Not a production or clinical system.