Release Notes: Ultrasound Ovarian Segmentation Sample Application#

Version 2026.3.0#

Release Date: TBD

This is the initial release of the Ultrasound Ovarian Segmentation sample application. It includes the following features:

  • Real-time, promptless ovarian tumor segmentation on 2D ultrasound using a SegFormer-B5 (DS2Net) model in the FP16 OpenVINO™ Toolkit (OpenVINO™) Intermediate Representation (IR) format.

  • Decoupled capture, inference, and display pipeline with an OpenCV overlay pop-up window and a heads-up display (HUD) showing FPS, latency, and GPU metrics.

  • Supported inputs: a video file, a USB webcam, or a folder of still images. Supported device: Intel® integrated GPU (iGPU), with a CPU fallback for development and sanity checks.

  • Built-in GPU utilization governor that enforces the KPI of less than 80% utilization on the busiest iGPU engine, with a per-run PASS or OVER verdict.

  • Model preparation backend (backend/bootstrap) that fine-tunes the model on the MMOTU-2D dataset, exports it to OpenVINO™ IR in FP16 precision, and verifies Dice and IoU scores. The application ships the recipe, not the model weights. Reference accuracy is Dice 0.8657 and IoU 0.7913 at resolution 256 on the MMOTU-2D validation split.

  • Optional sam model architecture (-Arch sam), selectable at runtime as an alternative to the default ds2net architecture. It uses a prompted pipeline that chains a YOLOv8n bounding-box detector, a Segment Anything Model (SAM) ViT-B encoder fine-tuned at resolution 256, and the stock SAM mask decoder. Reference accuracy is Dice 0.8207 and IoU 0.7394 on the MMOTU-2D validation split. Its model preparation code is in backend/bootstrap/sam. The SAM encoder and decoder path is licensed under Apache-2.0. The YOLOv8 detector (Ultralytics) is licensed under AGPL-3.0 and is used at build time only to produce the bounding-box IR. At runtime, the application loads only the exported OpenVINO™ IR and does not import ultralytics.