YOLO-US: A Unified Multi-Task Foundation Framework for Generalist Ultrasound Image Analysis

The inherent heterogeneity of Ultrasound (US) imaging impedes the development of generalizable medical models, whereas the clinical translation of Generalist Medical AI demands unified foundation models rather than task or organ-specific approaches. This study proposes YOLOUS, the official baseline for the FMC_UIA 2026 challenge, which enhances the YOLOv11 architecture with a custom C3_MViTBv3 backbone and lightweight MBTransformer blocks, achieving a balance between local texture extraction, global long-range dependency modeling, and computational efficiency. Shared multi-scale representations are dynamically routed to four decoupled, task-specific heads, which are optimized for segmentation, classification, detection, and biometric regression. Comprehensive evaluation across a heterogeneous benchmark consisting of 27 subtasks demonstrates state-of-the-art robustness, with a normalized IoU of 0.7955 for precise detetction and a normalized MRE of 40.63 for biometric measurement. Furthermore, in-depth analysis of segmentation performance under severe domain distribution shifts yields critical insights into the “heterogeneity gap,” validating YOLO-US as a scalable, mobile-friendly baseline for future ultrasound foundation models. The code will be available at https://github.com/LwpZoe/YOLO-US.