TRUST: Efficient Abdominal Trauma Recognition via Image-to-Ultrasound-Video Transfer Learning

2026-06-26

AI Summary

Abdominal ultrasound-based trauma identification suffers from substantial spatiotemporal and semantic variations due to differences in clinicians’ scanning habits, which hinders model generalization and interpretation efficiency. To address this challenge, this work proposes TRUST, a parameter-efficient image-to-video transfer learning framework that captures fine-grained spatiotemporal dynamics. The core innovations of TRUST include a cross-frequency collaborative adapter to enhance spatial feature extraction, a multi-granularity motion-aware module for robust temporal modeling, and a visual-query semantic aggregation mechanism to improve vision–language alignment. Experimental results demonstrate that TRUST achieves a 9.63% higher accuracy than the current state-of-the-art method on an internal abdominal trauma ultrasound dataset while maintaining superior computational efficiency.

Abstract

Abdominal ultrasound is indispensable for rapid, noninvasive trauma triage. However, interpreting the subtle dynamic cues embedded in continuous scanning is time-intensive and operator-dependent. Parameter-Efficient Image-to-Video Transfer Learning (PEIVTL), which efficiently adapts pre-trained image models to the video domain, notably through visual-textual alignment, offers a promising paradigm for ultrasound video analysis. Nevertheless, substantial spatiotemporal and semantic variations arising from physician-dependent scanning practices continue to limit the effectiveness and generalizability of this framework. We propose TRUST, a scan-aware PEIVTL framework that explicitly models fine-grained spatiotemporal variations to enable reliable ultrasound video understanding. First, we introduce a Cross-Frequency Collaborative Adapter (CFCA) that establishes mutual constraints between low- and high-frequency components, enhancing discriminative spatial feature extraction under heavy speckle corruption. Second, we design a Multi-Granularity Motion-Aware (MGMA) module that integrates local temporal convolutions with motion-prior-guided global self-attention, jointly capturing stable intra-view patterns and abrupt inter-view transitions to characterize complex scanning dynamics. Third, a Visual Query Semantic Aggregation (VQSA) module dynamically generates text prototypes conditioned on visual features, enabling adaptive visual-textual alignment robust to intra-class variability under diverse scanning conditions. Experiments on in-house ultrasound trauma datasets demonstrate that TRUST outperforms state-of-the-art methods by 9.63% with superior computational efficiency.

Problem

• abdominal trauma
• ultrasound video
• spatiotemporal variation
• operator dependency
• visual-textual alignment

Innovation

• Parameter-Efficient Transfer Learning
• Ultrasound Video Understanding
• Spatiotemporal Modeling
• Visual-Textual Alignment
• Motion-Aware Attention