Code and data
Our code releases live under one GitHub organisation, and our benchmark work is published openly so other groups can measure against it.
A hallucination evaluation benchmark for grounded medical multimodal LLMs: 67,000 visual-question-answering pairs with doctor-annotated anatomical segmentation masks, plus two evaluation protocols that test whether a model is reading the image or exploiting a textual shortcut. Published at IJCAI 2025 as Localizing Before Answering.
Much of our work uses South Australian health-service data: the RAPIDx AI trial cohort of about 14,000 patients across 12 hospitals, SA Pathology and SA Pharmacy records, and MyHealthRecord data under approved access. This data is governed and cannot be redistributed. Groups wanting to work with it do so through a formal collaboration, with ethics approval and a data-sharing agreement in place. That process usually sets the timeline for a project.
Each released repository carries the paper it belongs to in its description. Where a paper uses a public dataset such as MIMIC-CXR, the paediatric X-ray corpora or a public segmentation challenge, the repository documents which splits we used.
If a paper of ours has no repository yet and you need the code, please email us.
Artificial Intelligence in Medicine (Springer, 2022), the first comprehensive textbook on AI in medicine. Chapter on machine learning in medical imaging, senior author.
Intelligence-Based Cardiology and Cardiac Surgery (Elsevier, 2023), with two chapters, on medical visual question answering and on AI in echocardiography.
Co-author of the PRIME checklist (2021), cited internationally as a standard for evaluating cardiovascular imaging machine learning.