Abstract
Three-dimensional (3D) ultrasound (US) volumes offer rich, comprehensive anatomical information, yet they remain heavily underutilized in medical AI development. This underutilization stems from a lack of dedicated data-engineering frameworks capable of converting static, archived volumes into interactive, multi-task training data. Existing methods commonly rely on cross-modality synthesis of ultrasound images from CT or MRI data. This practice introduces simulation-to-real domain gaps and fails to preserve native ultrasound-specific acoustic artifacts. Furthermore, alternative approaches are generally restricted to isolated, single-task applications. To address these data-engineering limitations, we present Ultra GYM, a framework that transforms static 3D ultrasound archives into dynamic, interactive training environments. By operating directly on clinical 3D volumes, our platform preserves native acoustic properties and eliminates cross-modality domain shifts. Ultra GYM provides a modular data-engineering framework designed to easily extend to broader medical imaging pipelines. To showcase this multi-task scaling capability, we configure the framework to automatically generate specialized data structures for three highly distinct downstream tasks: (1) temporal video sweeps for standard plane detection, (2) exact six-degree-of-freedom tracking coordinates for 2D to 3D US reconstruction, and (3) an efficient reinforcement learning environment for autonomous Standard Plane detection. Across downstream evaluations, Models trained via the framework achieved 79.47\% overall accuracy in standard plane detection, an 84\% success rate in autonomous localization, and a local pose reconstruction error of about 0.12\,mm. These results demonstrate that Ultra GYM serves as an effective integrated framework, extracting diverse, multi-task training trajectories from limited clinical archives to accelerate medical AI deployment. The dataset, code, and videos will be made publicly available upon publication.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DEMI_023.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=vUqGMBPOtl
BibTex
@InProceedings{B.Rag_Ultra_MICCAISAT2026,
author = { B., Ragu AND Balaji, Sowjanya AND Burnwal, Returaj AND Ayyasamy, Shyam AND Seshadri, Suresh AND Lakshmanan, Manojkumar AND Sivaprakasam, Mohanasankar},
title = { { Ultra Gym: Transforming Clinical 3D Ultrasound Volumes into Dynamic, Multi-Task Training Data } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17276},
month = {pending},
page = {pending}
}