基于多尺度分形维数加权损失和注意力机制的鸡胚胎血管分割方法OA
Segmentation of chicken embryo vascular network based on multiscale fractal dimension weighted loss and attention mechanism
活体显微镜图像具有噪声高、动态性强的特点,而传统血管分割方法在应对这些挑战时经常会出现血管断连、预测错误的问题.为了解决这一问题,本研究设计了一种基于多尺度分形维数加权损失和注意力机制的深度学习模型,该模型通过分形维数加权损失和浅层注意力机制将血管网络的空间复杂度融入训练过程,同时通过多尺度结构和深层注意力机制保留了模型多层特征提取的优势和全局信息.在提出的私有数据集鸡胚血管网络图像数据上进行训练和测试,并通过血管断连数量和传统指标测试了其分割性能.结果显示,本研究提出的模型在整体上将 Dice 系数(Dice coefficient)从 0.896 提升至0.906,平均交并比(mean intersection of union,mIoU)从0.813提升至0.830,准确率(accu-racy)从0.885 提升至 0.938,ROC 曲线下面积(area under curve,AUC)从 0.914 提升至0.930,血管断连点的平均数量从 15.7 减少到 11.3.这表明模型在维持分割性能的同时,能够有效地减少血管断连的数量,提高了血管网络结构的连接性.
In vivo microscopy images are characterized by high noise and strong dynamics,and traditional vascular seg-mentation methods often encounter problems such as vascular discontinuity and prediction errors in addressing these challenges.To address this issue,this study designs a deep learning model based on multi-scale fractal dimension-weighted loss and attention mechanism.This model integrates the spatial complexity of the vascular network into the training process through fractal dimension-weighted loss and shallow attention mechanisms,while preserving the ad-vantages of multi-layer feature extraction and global information through multi-scale structures and deep attention mechanisms.The model is trained and tested on a proprietary dataset of chicken embryo vascular network images,and its segmentation performance is evaluated through the number of vascular discontinuities and traditional met-rics.The results show that the proposed model increases the Dice(Dice coefficient)index from 0.896 to 0.906,the mean intersection over union(mIoU)from0.813 to0.830,the accuracy from0.885 to0.938,the area under curve(AUC)from 0.914 to0.930,and reduces the average number of vascular discontinuity points from15.7 to 11.3.This indicates that the model effectively reduces the number of vascular discontinuities while maintaining seg-mentation performance,thereby improving the connectivity of the vascular network structure.
潘清;王硕;吕雪倩
浙江工业大学信息工程学院 杭州 310023浙江工业大学信息工程学院 杭州 310023浙江工业大学信息工程学院 杭州 310023
血管分割注意力机制分型维数多尺度损失函数活体显微镜成像
vessel segmentationattention mechanismfractal dimensionmultiscale loss functionin vivo microscopic imaging
《高技术通讯》 2026 (7)
701-712,12
国家自然科学基金(31870938,32371372)资助项目.
评论