基于全自动乳腺容积超声成像及三维深度学习的乳腺肿瘤分割与分类模型研究与应用

乳腺癌是女性最常见的恶性肿瘤，早期诊断对疾病预后至为关键。本项目基于课题组参与研发、国内首台全自动乳腺超声容积系统和三维卷积神经网络，开展乳腺BI-RADS多级分类和分割模型与理论的研究。项目针对医学数据缺乏、乳腺病灶的复杂性及识别与分割过程中需解决的不均匀性及个体差异等一系列问题，在完成肿瘤分割、分类金标准标签设计的基础上，基于全新无监督自编码器的域对抗神经网络预训练模型和肿瘤形态学描述的多模态特征提取技术，提取肿瘤深层次特征；并基于BI-RADS设计的多级分类胶囊网络，实现肿瘤分级判别；同时，基于正则化边界辅助监督的三维分割模型，初步分割出立体的乳腺肿瘤、腺体层、脂肪层与皮下组织层，并针对分割中图像结构和局部边界缺陷，研究基于监督机制的双向卷积循环网络，以有效解决分割边缘不完整性问题。该研究丰富与完善了三维卷积网络模型的理论与方法，为乳腺肿瘤人工智能辅助诊断的研究及应用探索了新的途径。

Breast cancer is a great threat to women's health. Medical imaging is of great significance for early diagnosis and clinical treatment of breast cancer. Based on the first automatic breast volume ultrasound imaging system in china, which is developed by the project team, and three-dimensional convolution neural network, the project will carry out the research on multi-level recognition, classification and segmentation model and theory of breast cancer. In view of complexity of breast lesions representation and lack of medical data, as well as the non-uniformity and individual differences that need to be solved in the recognition and segmentation process, the project constructs a new unsupervised self-encoder domain adversarial training neural network pretraining model and multi-feature extraction technique for the BI-RADS tumors morphological description after completing the design of tumor segmentation and classification gold standard label, so as to extract the deep-seated features of tumors; at the same time, using the boundary regularization supervision auxiliary 3D segmentation model, the preliminary segmentation of three-dimensional breast tumor, subcutaneouly tissue layer, lacteal gland tissue layer, fat behind glandular organ will be realized; aiming at the image structure and local boundary defects in segmentation mapping, the bi-directional convolution cycle network model based on the supervision mechanism is studied to effectively solve the problem of the incompleteness of segmentation edge. This research enriches and improves the theory and method of three-dimensional convolution network model and can explore a new way for the research and application of artificial intelligence aided diagnosis of breast cancer.