Task Aggregation Strategies for Ultrasound Foundation Models

A recent investigation published on arXiv (2603.18123) explores the reasons behind the underperformance of unified ultrasound foundation models compared to task-specific benchmarks. The researchers suggest that this decline in performance may result from task aggregation methods that overlook the interplay between the diversity of tasks and the scale of training data. They present M2DINO, a framework designed for multiple organs and tasks, which utilizes DINOv3 and incorporates task-conditioned Mixture-of-Experts blocks for flexible capacity management. The framework was tested across 27 ultrasound tasks, including segmentation, classification, detection, and regression, under three different paradigms: task-specific, clinically-grouped, and all-task. The findings provide guidelines for effectively learning heterogeneous ultrasound tasks together without sacrificing performance.