Overview
Ultrasound is the most widely deployed imaging modality worldwide, yet clinical AI remains fragmented into narrow, single-task models that fail when the device, operator, or anatomy changes. To address this gap, researchers present SonoCorpus and SonoBase β an open data resource and an interactive segmentation foundation model pretrained on it β as published on arXiv (2609.19230, cs.CV), submitted on 16 September 2026.
Authors: Chao Qin, Fahad Shahbaz Khan, Salman Khan, Sarim Ather, Siddiq Anwar, Rao Muhammad Anwer, and Shadab Khan.
Subject: Computer Vision and Pattern Recognition (cs.CV). The PDF includes the Supplementary Information.
SonoCorpus: A Unified Open Resource
SonoCorpus unifies:
- 456,963 images and 1,626,085 expert masks
- 53 public datasets spanning 24 clinical applications and 17 countries
This scale and diversity aim to overcome the fragmentation that limits generalization in ultrasound AI.
SonoBase: Foundation Model and Results
SonoBase is an interactive segmentation foundation model pretrained on SonoCorpus. Across fifteen evaluation datasets introducing new organs, devices, operators, and geographies, SonoBase:
- Outperforms SAM2, MedSAM2, and the concept-promptable MedSAM3 on every dataset
- Matches per-dataset specialist models trained on the same data
- On fully external data, exceeds the accuracy these baselines achieve on their own in-distribution benchmarks
Clinical Measurement Accuracy
- Ejection fraction (EF): derived segmentations fall within inter-observer variability, with a 6.63% error, and fewer misclassifications at the defibrillator-candidacy threshold than either promptable baseline (13% vs. 18β42%)
- Fetal head circumference: 1.81 mm error, below inter-observer variability
- Gestational age: 1.2 days error, below inter-observer variability
Robustness in the Field
Where a baseline fails outright β one in four test cases β SonoBase recovers a usable segmentation in 81% of them. This includes handheld probes operated by minimally trained users in two low- and middle-income countries (Sierra Leone and Tanzania).
Adaptability and Transferability
- Five labeled examples can help the model adapt to a new setting
- The identical training protocol transfers well to newer models such as SAM3, locating the advantage in ultrasound-specific pretraining rather than any single architecture
Reproducibility and Open Release
To ensure reproducibility and enable the community to build on SonoBase as a platform, the authors release:
- All checkpoints
- Optimizer states
- Data-split indices
- Deduplication hashes
- Starter code
Citation
Cite as: arXiv:2609.19230 [cs.CV] β https://arxiv.org/abs/2609.19230
via ArXiv CV