Transformer-Based Foundation Model for Universal Ultrasound Image Analysis

We propose a transformer-based foundation model for unified multi-task, multi-organ, multi-disease ultrasound image analysis, targeting 27 subtasks across segmentation, classification, detection, and landmark regression. The model follows hard parameter sharing: the encoder of a Swin Transformer pretrained on ImageNet is shared by all tasks, while each task uses its own prediction head. For segmentation and detection tasks, we use feature pyramid network decoders with separate segmentation and detection branches; for classification and regression tasks, we use encoder-level features with specific heads. Compared to other existing methods in offline experiments, the proposed method has the best performance. At the Foundation Model Challenge for Ultrasound Image Analysis (FMC UIA) challenge, our method achieved 8th place among 24th participating teams. The code is public at https://github.com/HJJ-D/Foundation-Model-Challenge-for-Ultrasound-Image-Analysis.