Towards Universal Ultrasound Analysis: Parameter-Efficient Foundation Model with Task-Aware Routing for Heterogeneous Multi-Task Learning

Developing a universal foundation model for ultrasound imaging is inherently challenging due to severe speckle noise, diverse acquisition protocols, and the semantic gap between heterogeneous clinical tasks. Traditional multi-task learning often suffers from negative transfer when simultaneously optimizing dense (i.e., segmentation, detection) and global (i.e., classification, regression) predictions. In this paper, we propose UltraTAR, a parameter-efficient Task-Aware Routing framework that won 1st Place in the Foundation Model Challenge for Ultrasound Image Analysis (FMC-UIA) 2026. Instead of full fine-tuning, we leverage a frozen Vision Foundation Model (VFM) adapted via Low-Rank Adaptation (LoRA) to preserve robust pre-trained representations while injecting lightweight task-specific knowledge. To overcome the architectural disconnect between single-scale foundation features and diverse task requirements, we design a task-aware dual-branch routing strategy, including a novel Feature Pyramid Adapter that constructs hierarchical FPN-compatible representations for dense tasks, and a CLS-guided feature fusion module that enriches global context for diagnostic tasks. Furthermore, an anchorfree CenterNet head and empirical loss scaling are integrated to harmonize optimization across 27 distinct subtasks. Evaluated on the large-scale FMC-UIA 2026 benchmark, our UltraTAR establishes a new state-of-the-art, outperforming the official baseline by large margins (e.g., +189% in Detection IoU and +19% in Segmentation DSC) and demonstrating exceptional versatility for universal ultrasound analysis.