基于U-Net的短轴心脏CTA左心室心肌自动分割系统的开发

Development of left ventricular myocardium automatic segmentation system from short axis CTA based on U-Net

目的开发并验证一种基于U-Net的短轴心脏计算机断层血管造影(CTA)左心室心肌(LVM)自动分割系统。方法选取2022年2至10月在温州市中心医院就诊的50例疑似或确诊心血管疾病患者CTA图像进行模型训练,将U-Net深化到8层,引入注意力机制和深监督机制,并加入残差连接。通过分割结果可视化观察、分割量化结果比较等分析本研究新开发方法与原型U-Net方法的左心室分割性能差异。结果可视化比较发现,无论分割小区域的左心室或分割组织对比度较差的图像,与原型U-Net方法相比,本研究新开发方法与标签具有更高的一致性;在可视化比较中,原型U-Net方法展现出欠分割、过分割等问题。与原型U-Net方法(0.938±0.144、0.941±0.144、0.961±0.058)相比,本研究新开发方法的Dice相似系数、精度、灵敏度分别为0.964±0.033、0.960±0.043、0.970±0.040,提示分割精度和鲁棒性更高。箱式图中可见两种方法均存在一些异常值。使用本研究新开发方法分割的3例患者左心室表面三维视图显示平均分割时间为13 s(即处理速度约为0.037 s/幅),提示分割质量和分割效率均较高。结论本研究新开发的基于U-Net的短轴心脏CTA LVM自动分割系统将注意力机制、深监督机制和残差连接集成到8层U-Net中,具有较高的分割精度、分割质量和分割效率。

Objective To develop and validate a short axis computed tomography angiography(CTA)left ventricular myocardium(LVM)automatic segmentation system based on U-Net.Methods The model training was conducted by using CTA images of 50 patients with suspected or confirmed cardiovascular disease in Wenzhou Central Hospital from February to October 2022.In the proposed method,the U-Net was deepened to 8 layers,attention and deep supervision mechanisms were introduced and residual connection was incorporated.The difference in left ventricular segmentation performance between the developed method and the prototype U-Net method was analyzed through visual observation of segmentation results and comparison of quantified segmentation results.Results In visual comparison,the developed method had higher consistency with the label than the prototype U-Net method,regardless of the small area of left ventricle segmentation or images with poor tissue contrast.Meanwhile,the prototype U-Net method exhibited problems such as under-segmentation and over-segmentation.Compared with the prototype U-Net method(0.938±0.144,0.941±0.144,0.961±0.058),the Dice similarity coefficient,accuracy and sensitivity of the developed method were 0.964±0.033,0.960±0.043,0.970±0.040,respectively,suggesting that the segmentation accuracy and robustness were higher.The box plot showed that both methods had some outliers.The three-dimensional view of the left ventricle surface of 3 patients using the newly developed method revealed an average segmentation duration of 13 s(i.e.the processing speed was 0.037 s/slice),indicating high segmentation quality and efficiency.Conclusion The developed LVM automatic segmentation system from short axis CTA based on U-Net,which integrates the attention mechanism,deep supervision mechanism and residual connection into 8-layer U-Net,has high segmentation accuracy,segmentation quality and segmentation efficiency.