TASL-Net: Tri-Attention Selective Learning Network for Intelligent Diagnosis of Bimodal Ultrasound Video

Chengqian Zhao, Zhao Yao, Zhaoyu Hu, Yuanxin Xie, Yafang Zhang, Yuanyuan Wang, Shuo Li, Jianhua Zhou, Jianqiao Zhou, Yin Wang, Jinhua Yu

Introduction

Medical diagnosis using bimodal ultrasound—which combines traditional gray-scale (GSUS) for anatomy and contrast-enhanced (CEUS) for blood flow—is a gold standard for many cancers. However, for deep learning models, these videos present a "black box" challenge: they are long, contain massive amounts of redundant data, and require a nuanced understanding of how blood perfusion changes over time.

Existing AI models often struggle to integrate the domain knowledge used by human experts, leading to high computational costs and lower clinical reliability. TASL-Net (Tri-Attention Selective Learning Network) addresses this by mimicking the diagnostic intuition of sonographers, embedding "tri-attention" (temporal, spatial, and bimodal) into a mutual transformer network.

The Tri-Attention Framework

Human sonographers don't watch the whole video with equal intensity; they have "diagnostic attention." They look for the exact moment contrast enters a lesion, focus on the differences between the tumor and surrounding tissue, and correlate the structural map of GSUS with the perfusion map of CEUS. TASL-Net breaks down this expertise into three core modules:

Overview of TASL-Net's Tri-Attention approach and the clinical challenges it addresses.
  1. TIC-based Video Selector (Temporal Attention): Instead of processing the entire video, this module uses Time-Intensity Curves (TIC) to automatically identify and "clip" the moments of rapid perfusion. This removes redundant frames and drastically improves efficiency.
  2. Earliest-Enhanced TIC Analysis (ETIC) (Spatial Attention): This module uses Structural Similarity (SSIM) variation to detect exactly where the contrast agent arrives first. It then differentiates between the lesion, healthy tissue, and surrounding structures (like the thoracic wall), ensuring the model compares the right areas.
  3. Mutual Convolutional Transformer (MCT) (Bimodal Attention): This hybrid architecture uses Convolutional Neural Networks (CNNs) to capture local textures and edges (ideal for GSUS) while Transformers model long-term temporal dependencies (ideal for CEUS). A feature-sharing unit allows these branches to communicate, bridging the semantic gap between the two imaging modes.

Experimental Results

The study validated TASL-Net on three distinct clinical datasets totaling 791 cases from multiple hospitals, covering lung, breast, and liver cancers. Notably, the researchers used real-world clinical data, including noise and variations, to ensure robustness.

TASL-Net outperformed five state-of-the-art (SOTA) methods, including TimeSformer and Video Swin Transformer (VST):

  • Diagnostic Accuracy: Achieved an average accuracy improvement of 6.43% over the next best method.
  • AUC Scores: Reached 0.86 for Lung, 0.86 for Breast, and 0.97 for Liver datasets.
  • Ablation Success: Removing any of the three "attentions" led to a significant drop in performance, confirming that the tri-attention approach is the primary driver of its success.

Conclusion and Future Work

TASL-Net demonstrates that better performance in medical AI comes from structured networks that follow human expert workflows rather than just increasing model depth. By filtering for "rapidly rising" perfusion moments and integrating bimodal synergy, the model becomes both more accurate and computationally efficient.

Future work will focus on increasing dataset diversity across more hospitals and developing "lightweight" versions of the network for point-of-care devices. The authors also aim to explore how this tri-attention strategy can be integrated with emerging large vision models.

Summary updated

BibTeX
@article{zhao2025tasl,
  author = {Chengqian Zhao and Zhao Yao and Zhaoyu Hu and Yuanxin Xie and Yafang Zhang and Yuanyuan Wang and Shuo Li and Jianhua Zhou and Jianqiao Zhou and Yin Wang and Jinhua Yu},
  title = {{TASL-Net: Tri-Attention Selective Learning Network for Intelligent Diagnosis of Bimodal Ultrasound Video}},
  journal = {Expert Systems with Applications},
  year = {2025},
  doi = {10.1016/j.eswa.2025.128355},
}
Figure