SEMC: Structure-Enhanced Mixture-of-Experts Contrastive Learning for Ultrasound Standard Plane Recognition

TL;DR

This work tackles ultrasound standard plane recognition, which suffers from high intra-class variance and subtle inter-class differences. It introduces SEMC, a structure-enhanced framework combining a Semantic-Structure Fusion Module (SSFM) with a Mixture-of-Experts Contrastive Recognition Module (MCRM) to fuse shallow structural cues with deep semantic features and to perform hierarchical contrastive learning across multiple expert branches. A high-quality LP2025 liver ultrasound dataset with six standard planes is released, and SEMC is evaluated against public benchmarks (FPUS23, CAMUS) and LP2025, achieving state-of-the-art performance across accuracy and F1-score. The method demonstrates robust, structure-aware representations and improved inter-class separability, with potential to enhance clinical workflow reliability in ultrasound plane identification.

Abstract

Ultrasound standard plane recognition is essential for clinical tasks such as disease screening, organ evaluation, and biometric measurement. However, existing methods fail to effectively exploit shallow structural information and struggle to capture fine-grained semantic differences through contrastive samples generated by image augmentations, ultimately resulting in suboptimal recognition of both structural and discriminative details in ultrasound standard planes. To address these issues, we propose SEMC, a novel Structure-Enhanced Mixture-of-Experts Contrastive learning framework that combines structure-aware feature fusion with expert-guided contrastive learning. Specifically, we first introduce a novel Semantic-Structure Fusion Module (SSFM) to exploit multi-scale structural information and enhance the model's ability to perceive fine-grained structural details by effectively aligning shallow and deep features. Then, a novel Mixture-of-Experts Contrastive Recognition Module (MCRM) is designed to perform hierarchical contrastive learning and classification across multi-level features using a mixture-of-experts (MoE) mechanism, further improving class separability and recognition performance. More importantly, we also curate a large-scale and meticulously annotated liver ultrasound dataset containing six standard planes. Extensive experimental results on our in-house dataset and two public datasets demonstrate that SEMC outperforms recent state-of-the-art methods across various metrics.

Figures (3)

• Figure 1: (a) Ultrasound standard planes exhibit large intra-class variance, where images from the same plane can appear markedly different, and small inter-class variance, where different planes often share highly similar visual patterns. (b) Previous methods mainly rely on deep semantic features, neglecting shallow structural cues. (c) In contrast, Our SEMC framework integrates the shallow structure via semantic-structure fusion and employs a MoE for hierarchical contrastive learning, which ca yields more discriminative and structure-aware representations.
• Figure 2: Architecture of the proposed SEMC framework. The framework first employs an MoE-based feature extractor to generate multi-level expert features from the input ultrasound image. These features are then aligned and enhanced through a Semantic-Structure Fusion Module (SSFM). The resulting representations are fed into the Mixture-of-Experts Contrastive Recognition Module (MCRM), which consists of two branches: a multi-class classification headserving as the primary task, and an MoE-based contrastive learning branch serving as an auxiliary task to further improve the primary task by refining the learned feature representations.
• Figure 3: Performance comparison for different values of the hyperparameter $\alpha$ on the LP2025 dataset.