FCT-Net:基于CNN与Transformer双分支并行融合的斑马鱼心脏图像分割网络模型

打开文本图片集

中图分类号：TP399 文献标志码：A 文章编号：1673-3851（2025）07-0571-09

引用格式：，.FCT-Net：基于CNN与Transformer双分支并行融合的斑马鱼心脏图像分割网络模型J.浙江理工大学学报（自然科学），2025，53（4）：571-579.

Abstract： To address the issue of insufficient accuracy in medical image segmentation caused by the blurred contours of zebrafish heart images，a novel network model called FCT-Net （fusion convolutiontransformer network）is proposed，which integrates CNN and Transformer. This model is based on the classic encoder-decoder architecture and a dual-branch parallel feature fusion module is constructed. Specifically，the CNN branch is utilized to extract local tissue features.To overcome the limitation of a single convolutional kernel in covering multi-scale features，a multi-scale feature fusion mechanism is introduced within the convolutional module，and a multi-receptive field feature pyramid is built to enhance the representation capability of edge details. The Transformer branch is employed to capture long-range global contextual dependencies，achieving effective fusion of local features and global semantics. Experimental results demonstrate that FCT-Net improves the accuracy by 5.8% compared to the baseline U-Net model in the task of zebrafish heart image segmentation，effectively enhancing the precision of heart contour segmentation. With its high-precision zebrafish heart segmentation capability，this model can provide relatively reliable algorithmic support for the subsequent drug screening studies based on the morphological characteristics of the zebrafish heart.

Key words： zebrafish; heart image;CNN; Transformer; multi-scale feature fusion

0 引言

近年来，斑马鱼模型在心脏功能研究领域取得了较为明显的进展，通过分析斑马鱼心脏形态可以为药物筛选和疾病机理探索等生物医学研究提供数据支持[1-2]，但现有的斑马鱼心脏形态分析方法仍面临不少挑战。（剩余13289字）