EndoSeg: What We Built, What We Didn't, and Why the Overlay Lies a Little

Short summary

EndoSeg is an ovarian-lesion ultrasound segmenter fine-tuned on Nebius Serverless, exported to ONNX, and run entirely in-browser so scans never leave the device. The author is transparent that the model outputs only a binary lesion-vs-background mask — not a diagnosis — while UI labels like 'chocolate cyst' are driven by JS shape heuristics, not learned classifiers. The post serves as a cautionary guide for ML practitioners: don't let product UI narrate capabilities the model was never trained to produce.

• •Binary lesion segmenter fine-tuned on MMOTU ultrasound data, deployed via ONNX Runtime Web for on-device inference
• •UI labels (e.g. 'endometrioma') come from JS shape heuristics, not the neural network — a deliberate scoping gap the author discloses
• •Privacy-by-design: model weights cached locally, images stay on-device unless user opts into cloud comparison

Generated with AI, which can make mistakes.