基于nnU-Net的乳腺DCE-MR图像中乳房和腺体自动分割

在乳腺动态增强磁共振（DCE-MR）图像中，乳房分割和腺体分割是进行乳腺癌风险评估的关键步骤.为实现在三维脂肪抑制乳腺DCE-MR图像中乳房和腺体的自动分割，本文提出一种基于nnU-Net的自动分割模型，利用U-Net分层学习图像特征的优势，融合深层特征与浅层特征，得到乳房分割和腺体分割结果.同时，基于nnU-Net策略，所使用的模型能根据图像参数自动进行预处理和数据扩增，并动态调整网络结构和参数配置.实验结果表明，在具有多样化参数的三维脂肪抑制乳腺DCE-MR图像数据集上，该模型能准确、有效地实现乳房和腺体分割，平均Dice相似系数分别达到0.969±0.007和0.893±0.054.

Automatic Segmentation of Breast and Fibroglandular Tissues in DCE-MR Images Based on nnU-Net

Segmentation of whole breast and fibroglandular tissue (FGT) is an important task for quantitative analysis of breast cancer risk in dynamic contrast enhanced magnetic resonance (DCE-MR) images. In this study, an automated segmentation model based on nnU-Net is proposed to segment the whole breast and FGT in 3D fat-suppressed breast DCE-MR images, taking the advantages of hierarchical image features learning, as well as the fusion of deep features and shallow features. The model could automatically perform preprocessing, data augmentation and dynamic adaptation of network configurations with respect to different imaging parameters. Experimental results show that the method could accurately and efficiently segment the whole breast and FGT in the collected dataset of 3D fat-suppressed breast DCE-MR images with variable imaging characteristics, achieving the average Dice similarity coefficients 0.969±0.007 and 0.893±0.054, respectively, for breast and FGT segmentation.