On the public dissemination and open sourcing of ultrasound resources, datasets and deep learning models.
Authors
Affiliations (3)
Affiliations (3)
- Institute of Biomedical Engineering, Department of Engineering Science, Roosevelt Drive, Old Road Campus Research Building, Headington, Oxford, Oxfordshire, United Kingdom. [email protected].
- Department of Computer Science, Khalifa University, Shakhbout Bin Sultan Street, Hadbat Al Za'faranah, Zone 1, Abu Dhabi, Abu Dhabi, United Arab Emirates. [email protected].
- Institute of Biomedical Engineering, Department of Engineering Science, Roosevelt Drive, Old Road Campus Research Building, Headington, Oxford, Oxfordshire, United Kingdom.
Abstract
Ultrasound data is relatively under-utilized in machine learning applied to medical imaging research when compared to other imaging modalities. Towards rectifying this, this paper (and the associated webpage) catalogs and assesses the usability of publicly available ultrasound datasets and models. Datasets were categorized and ranked using an original dataset quality score, SonoDQS. The models were scored using our model quality score, SonoMQS. We identified 72 public ultrasound datasets covering different anatomies and collected in different parts of the world. We identified 56 open-source models trained on ultrasound data. Most open-source models were trained on datasets that are or were made publicly available. A plurality of the datasets are of similar quality, corresponding to bronze (fifth tier) in the SonoDQS ranking. There are a few publicly available datasets of fetal content (5) and prostate anatomy (4) in spite of the wide use of ultrasound in these clinical areas, acknowledging a notable gap.