Attention-guided model for mitral regurgitation analysis based on multi-task learning

• Published in: Biomedical signal processing and control 2025-03, Vol.101, p.107169, Article 107169
• Main Authors: Wu, Jing, Ge, Zhenyi, Huang, Helin, Wang, Hairui, Li, Nan, Hu, Chunqiang, Pan, Cuizhen, Wu, Xiaomei
• Format: Article
• Language: English
• Subjects: Attention mechanism; Echocardiography; Mitral regurgitation; Multi-task learning; Self-supervised learning
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2024.107169

For automated analysis of mitral valve regurgitation in non-Doppler-based 2D echocardiography, there is limited work on combining quantitative tasks for cardiac targets, such as semantic segmentation and motion tracking, with qualitative detection and etiological classification of mitral regurgitation, thus overlooking the common features that exist among these tasks. Therefore, we proposed a multi-task learning model called Attention-guided ResUNet-MTL (abbreviated as ARUNet-MTL), to address the task from a holistic view. Specifically, the model was built on the U-shaped architecture and emphasized the importance of inherent correlation among tasks by allowing all three tasks to share the encoder structure. Meanwhile, an attention mechanism called MSA was incorporated to improve the temporal continuity of segmentation image sequences by leveraging the bidirectional deformation field information achieved in motion tracking tasks. Besides, during the training phase of the model, the loss function was designed to assimilate two key aspects: the fidelity of cardiac anatomical structures in segmentation and motion tracking results, and the enforcement smooth transitions and coherence between consecutive frames in the sequence. Through 5-fold cross validation, the accuracy for mitral regurgitation etiological classification was 0.8946 and 0.9179 at the video level and subject level, respectively, while the macro-F1 score was 0.8976 and 0.9176, respectively. The segmentation results for the left atrium, left ventricle, and mitral valve yielded Dice coefficients of 0.9438, 0.9157, and 0.7951. Additionally, validation experiments were performed on two public datasets to verify the robustness of the model’s segmentation and motion tracking branches. •The model processed 2DE for MR classification, segmentation, and motion tracking.•The model excels in all tasks by using shared features and attention mechanism.•The study supports rapid screening of general MR types with promising prospects.