基于上下文感知的轻量化FAST腹腔积液与解剖结构联合识别-《中国医学物理学杂志》

文章信息/Info

Title:: Context-aware lightweight FAST for joint recognition of ascites and anatomical structures

作者:: 荆娟1; 唐嘉杰2; 肖晶晶2; 王伟1; 姜小明1; 1.重庆邮电大学生命健康信息科学与工程学院，重庆 400065； 2.陆军军医大学第二附属医院生物医学大数据与人工智能中心，重庆 400037

Author(s):: JING Juan1; TANG Jiajie2; XIAO Jingjing2; WANG Wei1; JIANG Xiaoming1; 1. School of Life Health Information Science and Engineering, Chongqing University of Posts and Telecommunications, Chongqing 400065, China 2. Bio-Med Informatics Research Centre & Clinical Research Centre, Xinqiao Hospital, Army Medical University, Chongqing 40037, China

Keywords:: Keywords: focused assessment with sonography for trauma abdominal ultrasonography context-aware multi-task learning lightweight

摘要:: 针对创伤重点超声评估（FAST）技术高度依赖操作者经验及现有深度学习模型多任务协同差、计算效率低、病灶识别定位能力不足等瓶颈，提出一种基于上下文感知的轻量化多任务模型（CALM-FAST），通过特征共享机制与任务特异性表征有机结合，实现腹腔积液区域分割与解剖结构检测的联合优化。研究在主干网络引入上下文感知轻量模块，结合快速卷积和特征融合策略，在保证全局特征表征能力的同时将计算复杂度降低至6.90 GFLOPs；提出多任务联合损失函数以缓解FAST数据不平衡问题；构建包含693张超声影像真实世界数据集，采用分层迁移学习策略增强模型泛化能力。结果表明，本文方法在解剖结构检测任务mAP@0.5和mAP@0.5:0.95分别达到93.0%和64.5%；腹腔积液分割任务IoU指标提升至25.9%，较基线模型提升9.0%。

Abstract:: Abstract: Focused assessment with sonography for trauma (FAST) technology highly depends on operator experience, and existing deep learning-based models encounter key technical challenges including inadequate multi-task collaboration, low computational efficiency, and insufficient ability to identify and locate lesions. To address the above issues, a context-aware lightweight multi-task model (CALM-FAST) is proposed, which combines feature sharing mechanism with task-specific representation to achieve joint optimization of peritoneal effusion segmentation and anatomical structure detection. The proposed model introduces a context-aware lightweight module into the backbone network, and integrates fast convolution and feature fusion strategies, which reduces the computational complexity to 6.90 GFLOPs while preserving global feature representation capability. Furthermore, a multi-task joint loss function is used to mitigate the data imbalance issue in FAST. In addition, a real-world dataset containing 693 ultrasound images is constructed, and a hierarchical transfer learning strategy is adopted to enhance the generalization performance. Results show that the proposed method achieves a mAP@0.5 of 93.0% and a mAP@0.5:0.95 of 64.5% in the anatomical structure detection task. For the peritoneal effusion segmentation, the IoU is increased to 25.9%, which is 9.0% higher than the baseline model.