摘要： 斑马鱼是心脏疾病研究的常用模式生物，其幼体心脏呈透明状态，能够在显微镜下直接观察，目前尚未有成熟有效的算法自动识别斑马鱼心脏。为提高斑马鱼心室图像自动分割的准确率和实时效率，解决心室区域弱边界、特征提取能力不足及帧间相关性细节利用不充分等问题，本研究提出了一种基于边界引导特征库的斑马鱼心室图像自动分割方法EGAFB。通过嵌入矩形自校准模块改进编码器，提取全局上下文信息，增强心室特征提取；同时，引入边界引导注意力机制，引导模型关注更多的边界细节信息，强化算法的心室边界识别能力；引入均方误差损失和分类置信度损失函数，优化模型的局部细化分割机制，提高斑马鱼心室识别准确率。结果显示：EGAFB方法的平均交并比mIoU达94.7%。与现有方法Unet相比，mIoU提升4.6%，推理时间减少22.9%；与原模型相比，mIoU提升1.3%，推理时间减少6.4%。研究表明本方法在准确率和实时分割效率方面具有显著优势，为斑马鱼心室图像自动分割提供有效的解决方案，同时为斑马鱼心脏疾病模型研究提供更高效的技术支持。

关键词: 斑马鱼心室识别, 图像语义分割, 特征提取, EGAFB, 边界注意力机制

Abstract: Zebrafish is a commonly used model organism for heart disease research, and its larval cardiac is transparent and able to be directly observed under the microscope, and no mature and effective algorithm has yet been developed to automatically identify the zebrafish cardiac. In order to improve the accuracy and real-time efficiency of automatic segmentation of zebrafish ventricular images, and to solve the problem of weak boundaries of the ventricular region, insufficient feature extraction capability, and insufficient utilization of inter-frame correlation details, this thesis proposes EGAFB, a method for zebrafish ventricular images segmentation based on an edge-guided adaptive feature bank. The encoder is improved by embedding a rectangular self-calibration module, which extracts the global context information and enhances the ventricular feature extraction; meanwhile, the edge-guided attention mechanism is introduced to direct the model to focus on more boundary detail information, which strengthens the ventricular boundary identification ability; mean square error loss and classification confidence loss functions are introduced to optimise the local refinement segmentation mechanism, which improves the zebrafish ventricle recognition accuracy. The results show that the mean Intersection over Union (mIoU) of the EGAFB reaches 94.7%. Compared with the existing method Unet, the mIoU is improved by 4.6% and the inference time is reduced by 22.9%; compared with the original model, the mIoU is improved by 1.3% and the inference time is reduced by 6.4%. This thesis shows that the EGAFB method has high accuracy and real-time segmentation efficiency, which provides an effective solution for automatic segmentation of zebrafish ventricular images, as well as some more efficient technical support for the research of zebrafish cardiac disease models. 

Key words: Zebrafish ventricle recognition, image semantic segmentation, feature extraction, EGAFB, edge-guided attention mechanism