EndoSeg: A Gynecological Ultrasound Segmentation Tool

Browser-native gynecological (ovarian-lesion) ultrasound segmentation built for the
Nebius Serverless AI Builders Challenge — Healthcare & Life Sciences track.

Fine-tuned in the cloud. Inferred privately in your browser — your scan never leaves your laptop.

Overview

EndoSeg is a complete cloud-to-browser medical image segmentation pipeline. A compact
2D U-Net is fine-tuned on the MMOTU ovarian-tumor ultrasound dataset using Nebius
Serverless Jobs — one job for preprocessing, one for GPU training, and one for ONNX
export. The exported model is quantized to under 50 MB and shipped inside a static
webpage where all inference runs locally in the browser using ONNX Runtime Web and
WebGPU. No server, no upload, no waiting.

The privacy-by-design angle is the distinguishing feature. Medical images are sensitive
by nature; most cloud demos require users to upload their scans. EndoSeg inverts this
model: the trained weights are fetched once and cached, then every inference happens
entirely on the user's device using client-side JavaScript. A second, heavier
foundation-model fine-tune lives on a Nebius Endpoint for a "compare to cloud"
feature — but this path is opt-in, routed through a token-hiding proxy, and clearly
labelled so users always know where their data goes.

Architecture

Two-model split: a compact U-Net runs entirely in the browser via ONNX Runtime Web
(quantized, ≤ 50 MB, 256×256, WebGPU with WASM fallback). A heavier foundation-model
fine-tune (UltraSam / MedSAM2) lives on the Nebius Endpoint for the optional
"compare to cloud" feature — higher accuracy at the cost of a round-trip through the
token-hiding proxy.