Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Baseline Method of the Foundation Model Challenge for Ultrasound Image Analysis

Bo Deng, Yitong Tang, Jiake Li, Yuxin Huang, Li Wang, Yu Zhang, Yufei Zhan, Hua Lu, Xiaoshen Zhang, Jieyun Bai

TL;DR

The paper tackles the challenge of building generalizable foundation models for ultrasound by introducing FM_UIA 2026, a large-scale, multi-task benchmark spanning 27 subtasks in segmentation, classification, detection, and regression. It proposes a unified Multi-Head Multi-Task Learning baseline based on an EfficientNet-B4 encoder and a Feature Pyramid Network, with a task-specific routing scheme that differentiates global and dense tasks within a single network. Results on the official validation set show robust performance on global tasks and reveal limitations in precise localization and measurement tasks, highlighting the need for improved multi-scale processing and localization modules. The work provides a strong, extensible baseline and publicly available code and data to accelerate ultrasound foundation model research.

Abstract

Ultrasound (US) imaging exhibits substantial heterogeneity across anatomical structures and acquisition protocols, posing significant challenges to the development of generalizable analysis models. Most existing methods are task-specific, limiting their suitability as clinically deployable foundation models. To address this limitation, the Foundation Model Challenge for Ultrasound Image Analysis (FM\_UIA~2026) introduces a large-scale multi-task benchmark comprising 27 subtasks across segmentation, classification, detection, and regression. In this paper, we present the official baseline for FM\_UIA~2026 based on a unified Multi-Head Multi-Task Learning (MH-MTL) framework that supports all tasks within a single shared network. The model employs an ImageNet-pretrained EfficientNet--B4 backbone for robust feature extraction, combined with a Feature Pyramid Network (FPN) to capture multi-scale contextual information. A task-specific routing strategy enables global tasks to leverage high-level semantic features, while dense prediction tasks exploit spatially detailed FPN representations. Training incorporates a composite loss with task-adaptive learning rate scaling and a cosine annealing schedule. Validation results demonstrate the feasibility and robustness of this unified design, establishing a strong and extensible baseline for ultrasound foundation model research. The code and dataset are publicly available at \href{https://github.com/lijiake2408/Foundation-Model-Challenge-for-Ultrasound-Image-Analysis}{GitHub}.

Baseline Method of the Foundation Model Challenge for Ultrasound Image Analysis

TL;DR

The paper tackles the challenge of building generalizable foundation models for ultrasound by introducing FM_UIA 2026, a large-scale, multi-task benchmark spanning 27 subtasks in segmentation, classification, detection, and regression. It proposes a unified Multi-Head Multi-Task Learning baseline based on an EfficientNet-B4 encoder and a Feature Pyramid Network, with a task-specific routing scheme that differentiates global and dense tasks within a single network. Results on the official validation set show robust performance on global tasks and reveal limitations in precise localization and measurement tasks, highlighting the need for improved multi-scale processing and localization modules. The work provides a strong, extensible baseline and publicly available code and data to accelerate ultrasound foundation model research.

Abstract

Ultrasound (US) imaging exhibits substantial heterogeneity across anatomical structures and acquisition protocols, posing significant challenges to the development of generalizable analysis models. Most existing methods are task-specific, limiting their suitability as clinically deployable foundation models. To address this limitation, the Foundation Model Challenge for Ultrasound Image Analysis (FM\_UIA~2026) introduces a large-scale multi-task benchmark comprising 27 subtasks across segmentation, classification, detection, and regression. In this paper, we present the official baseline for FM\_UIA~2026 based on a unified Multi-Head Multi-Task Learning (MH-MTL) framework that supports all tasks within a single shared network. The model employs an ImageNet-pretrained EfficientNet--B4 backbone for robust feature extraction, combined with a Feature Pyramid Network (FPN) to capture multi-scale contextual information. A task-specific routing strategy enables global tasks to leverage high-level semantic features, while dense prediction tasks exploit spatially detailed FPN representations. Training incorporates a composite loss with task-adaptive learning rate scaling and a cosine annealing schedule. Validation results demonstrate the feasibility and robustness of this unified design, establishing a strong and extensible baseline for ultrasound foundation model research. The code and dataset are publicly available at \href{https://github.com/lijiake2408/Foundation-Model-Challenge-for-Ultrasound-Image-Analysis}{GitHub}.
Paper Structure (20 sections, 6 equations, 1 figure, 1 table)

This paper contains 20 sections, 6 equations, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. METHOD
  3. Unified Network Architecture
  4. Shared Encoder
  5. Feature Pyramid Decoder
  6. Task-Specific Routing and Heads
  7. Global Tasks (Classification & Regression)
  8. Dense Tasks (Segmentation & Detection)
  9. Objective Functions
  10. Segmentation Loss
  11. Classification Loss
  12. Regression Loss
  13. Detection Loss
  14. EXPERIMENTS AND RESULTS
  15. Datasets
  16. ...and 5 more sections

Figures (1)

  • Figure 1: Overview of the proposed baseline framework. The model utilizes an EfficientNet-B4 backbone for feature extraction. Conditioned on the input task identifier, a task-specific routing mechanism directs the high-level semantic features ($C_5$) to the Global Average Pooling (GAP) module for global tasks (Classification, Regression) or aggregates multi-scale features via the Feature Pyramid Network (FPN) decoder to generate high-resolution feature maps $P_{out}$ for dense tasks (Segmentation, Detection).