UniUltra: A Unified Foundation Model for Universal Ultrasound Analysis

Ultrasound (US) imaging exhibits substantial heterogeneity across anatomical structures, posing significant challenges to the development of generalizable foundation models. In this paper, we present UniUltra, a unified Transformer-based framework designed for the Foundation Model Challenge for Ultrasound Image Analysis (FMC_UIA) 2026. To overcome the resolution bottlenecks and long-range dependency limitations observed in convolutional baselines, our architecture utilizes a Swin V2 Tiny backbone with Log-spaced Continuous Position Bias (Log-CPB) and supports an increased input resolution of$384 \times 384$. We introduce a Single Query Guided Head for classification to adaptively fuse semantic and pixel-level features, and integrate a Differentiable Spatial Numerical Transform (DSNT) layer to achieve subpixel accuracy in biometric regression. Quantitative evaluation across 27 heterogeneous subtasks demonstrates the robustness of our approach, yielding an average Dice Similarity Coefficient (DSC) of 0.8910 in segmentation and an Area Under the Curve (AUC) of 0.9313 in classification. Furthermore, the model significantly improves localization precision, attaining a mean Intersection over Union (IoU) of 0.7831 in detection and a Mean Radial Error (MRE) of 18.80 in regression. These results validate the efficacy of our Transformer-based multi-task approach in complex clinical ultrasound analysis.