用卷积神经网络U-Net分割对比增强锥光束乳腺CT数据

Segmentation of contrast-enhanced cone-beam breast CT data using U-Net convolutional neural networks

乳腺肿瘤是一种常见的恶性肿瘤,其临床诊断不但费时费力,还容易出现误诊.本文旨在建立一个基于乳腺数据自动分割的乳腺肿瘤计算机辅助诊断模型,提高临床诊断的速度和准确率.为了用卷积神经网络U-Net模型分割对比增强锥光束乳腺计算机断层扫描(Contrast-Enhanced Cone-Beam Breast CT,CE-CBBCT)数据,本文首先沿冠状面将3维数据转换成2维切片,通过数据默认的窗口对其进行归一化处理.实验结果显示,使用U-Net卷积神经网络对数据进行分割,Dice系数和IoU(IntersectionoverUnion)分别为0.7920和0.6962.然后,本文用不同骨干网络(即各种深度学习分类网络)去替换U-Net的编码器,再次进行分割并对比不同特征提取对分割性能的影响,发现旋转增广方式可以提升各分割网络的性能.其中,基于ResNet152的U形分割网络的性能最好,Dice系数和IoU分别达到0.8410和0.7576.随后,本文又在所有模型中选取5个性能最好的模型组成一个集成模型,重复分割实验,发现此模型有最佳分割性能,平均Dice系数和IoU可达0.8463和0.7676,性能显著提升.值得指出的是,在处理数据时本文仅使用数据默认的窗口,降低了对人工的依赖.

Breast cancer is a common malignant disease.Early diagnosis is crucial to reduce the mortality rate and improve the treatment outcome of breast cancer.However,the clinical diagnosis of breast cancer is both time-consuming and labor-intensive with misdiagnosis risk.This paper aims to improve the diagnostic speed and accuracy of breast cancer,a computer-aided diagnostic model is established for the automatic segmenta‐tion of breast CT data.We take the high-resolution 3D contrast-enhanced cone-beam breast computed tomog‐raphy breast data as a dataset and segment it by using different convolution-based neural network U-Nets.Firstly,we convert the 3D dataset into 2D slices along the coronal plane,normalize them by the default win‐dow provided by the image data dicom.Then we implement the segmentation.It is shown that,when the original U-Net network is used,the Dice coefficient and IoU(Intersection over Union)are 0.7920 and 0.6962,respectively.Then,by augmenting the training samples through rotation,i.e.,rotating each train‐ing image every 10 degrees to increase the training samples,we implement the segmentation again.It is shown that the Dice coefficient and IoU are boosted to 0.8261 and 0.7418,respectively.Then we replace the encoder part of U-Net by different backbone networks,i.e.,various commonly-used deep learning classifica‐tion networks,and implement the segmentation again to compare the effect of different extraction features on the segmentation performance.It is shown that the U-shaped segmentation network based on ResNet152 has the best performance with the Dice coefficient 0.8367 and IoU 0.7502.Notably,the segmentation perfor‐mance of all networks is improved by the same rotation augmentation.However,it is still the ResNet152-based U-shaped segmentation network keeps the best performance with the Dice coefficient 0.8410 and IoU 0.7576.Finally,we choose the top five models from these models and build an integrated model.It is shown that,compared to the baseline network U-Net,the integrated model significantly improves the segmentation performance with the average Dice coefficient 0.8463 and IoU 0.7676.It is worthy of pointing out that all segmentation models use the default window provided by the image for data normalization and reduce the de‐pendence on manual work.