基于多尺度集成的冠状动脉分割模型

Coronary artery segmentation model based on multiscale intergration

针对传统冠状动脉分割中需要人为干预且效率低,以及现有深度学习分割方法准确率低的问题,本研究提出一种基于多尺度集成的分割模型。该模型设计了一种新的由粗到细的分割框架,通过结合全尺度的粗分割与局部多尺度的细分割,进一步提升分割的准确率。实验结果表明,在Dice相似性系数上可达到82.96%,优于其他常规的深度学习方法。该模型也为其他管状器官的分割提供了新的思路。

To address the problem that human intervention and low efficiency in traditional segmentation methods and improve accuracy in deep learning segmentation methods, we proposed a coronary artery segmentation model based on multi-scale integration. This model mainly used a coarse-to-fine segmentation framework, and further improved the accuracy of the segmentation by combining coarse segmentation at the global scale with fine segmentation at multiple local scales. Experimental results showed that it achieved 82.96% in Dice metrics, and was better than other conventional deep learning methods.It also provides new ideas for segmentation of other tubular organs.