SonoBase

Paper: arXiv:2609.19230 · Code: github.com/AlfredQin/sonobase · Data manifest: 10.5281/zenodo.22770825 · Demo: Hugging Face Space · Website: alfredqin.github.io/sonobase-web

SonoBase is an interactive ultrasound segmentation foundation model. It adapts SAM2 with an image-pyramid hybrid encoder (a Hiera-B transformer branch at 256 px and two ConvNeXt branches at 512 / 1024 px, joined by cross-branch attention) and is pretrained on SonoCorpus, a corpus of 53 public ultrasound datasets (456,963 images/frames, 1,626,085 masks, 24 clinical applications).

Files

File	What it is
sonobase_hiera_b_conv_s_conv_t.pt	Model weights (state dict under model, plus _provenance), 706 MB. Load with the sonobase recipe in the code release.
checkpoint_dcp/epoch_19_step_27539/	Full PyTorch distributed checkpoint of the production run: model/, optim/ (AdamW optimizer states), rng/, scaler.pt, step_scheduler.pt, config.yaml, losses.json. Resume or fine-tune from here.
config.yaml	Resolved training configuration of the production run.
training.jsonl, validation.jsonl	Per-step training log and per-epoch validation log.

Production run: hiera_b_conv_s_conv_t, 20 epochs on the 38 pretraining + 8 benchmark-train datasets of SonoCorpus with the leak-free 38_pt_8_bm_7_ext_v2 splits, global batch 16, cosine LR 5e-5 to 5e-6 (3e-5 to 3e-6 for the encoder), AdamW, weight decay 0.1, layer decay 0.9, resolution 1024, bf16, 2 nodes x 8 H200. Checkpoint epoch_19_step_27539 is both the last and the lowest-validation-loss epoch (val mIoU 0.8385).

Usage

See https://github.com/AlfredQin/sonobase (Phase 3/4 scripts: src/scripts/pretrain/, src/scripts/benchmarks/). The .pt file is the DCP checkpoint converted with the repository's DCP-to-PT utility; the benchmark, few-shot and analysis recipes consume it directly.

Interactive demo

Point and box prompts on images, and prompt-once propagation on short clips, with this checkpoint: https://huggingface.co/spaces/AlfredQin/sonobase-demo (research use only, not a medical device).

Data

The SonoCorpus manifest (dataset sources and licences, split lists, per-file checksums, per-unit metadata and evaluation records) is published on Zenodo: https://doi.org/10.5281/zenodo.22770825 (CC BY 4.0). No images or masks are re-hosted; datasets are obtained from their sources under their own terms. Several constituent datasets are non-commercial, which constrains downstream use of these weights accordingly.

Citation

@article{sonobase2026,
  title         = {Open ultrasound foundation model for robust segmentation and clinical measurement across heterogeneous settings},
  author        = {Qin, Chao and Khan, Fahad Shahbaz and Khan, Salman and Ather, Sarim and
                   Anwar, Siddiq and Anwer, Rao Muhammad and Khan, Shadab},
  journal       = {arXiv preprint arXiv:2609.19230},
  year          = {2026},
  eprint        = {2609.19230},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2609.19230}
}

Please also cite the original publication of every constituent dataset you use.

Licence

The SonoBase model weights and optimizer states in this repository are released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) licence: several constituent datasets of SonoCorpus are non-commercial, so the weights are non-commercial too. The code is MIT-licensed (https://github.com/AlfredQin/sonobase) and the SonoCorpus manifest is CC BY 4.0 (Zenodo). SonoBase is a research tool and not a medical device.