基于全卷积网络的乳腺肿瘤动态增强磁共振图像分割

Segmentation of Breast Tumors Based on Fully Convolutional Network and Dynamic Contrast Enhanced Magnetic Resonance Image

准确可靠的乳腺肿瘤分割是乳腺癌诊断、治疗、预后评估的关键.针对现有的基于动态增强磁共振成像(DCE-MRI)的乳腺肿瘤分割方法易遗漏小目标肿瘤等不足,本文提出了一种基于全卷积网络的可靠高效的乳腺肿瘤DCE-MRI图像分割方法.首先,对乳腺DCE-MRI数据进行预处理后,截取128*128大小的图像块,并以肿瘤区域像素数为依据将数据分为两个子数据集;其次,利用数据集训练CBP5-Net得到分类模型;然后,利用两个子数据集分别训练RAU-Net得到两个分割模型;最后,将测试集数据送到网络输入端,并对网络输出结果进行后处理,得到最终的乳腺肿瘤分割结果.利用本文提出的方法得到的Dice系数、敏感性、特异性和交并比(IoU)分别达到了0.9388、0.9523、0.9985和0.8768,说明利用本文方法能够有效、精确地分割乳腺肿瘤DCE-MRI图像.

Accurate and reliable breast tumor segmentation is essential for the diagnosis,treatment and prognosis of breast cancer.To address the shortcomings of existing dynamic contrast enhanced-magnetic resonance imaging(DCE-MRI)-based breast tumor segmentation methods,which tend to miss small tumors,we proposed a more reliable and efficient segmentation method for breast tumors in DCE-MRI based on a fully convolutional network(FCN).Firstly,the breast DCE-MRI data was preprocessed,followed by intercepting the image blocks of 128*128,and dividing the dataset into two sub-datasets according to the number of pixels in the tumor region.Secondly,the whole set was used to train CBP5-Net to obtain a classification model.Then,two sub-datasets were used to train RAU-Net to get two segmentation models.Finally,the test set was entered into the network,and the network outputs were post-processed to obtain the final segmentation results.The Dice coefficient,sensitivity,specificity and intersection over union(IoU)index of the method proposed in this paper reached 0.9388,0.9523,0.9985 and 0.8768,respectively.It proves that the proposed method can be used to segment DCE-MRI breast tumors effectively and accurately.