乳腺癌超声图像辅助诊断技术对于乳腺癌超声筛查、辅助发现肿瘤、提高诊断的准确率有重要临床意义。项目拟针对乳腺癌超声图像辅助诊断关键技术进行深入研究，研究基于多维度特征的乳腺癌超声图像分割、分类新方法。主要包括：（1）乳腺超声图像肿瘤及相关区域自动分割新方法。研究基于低层特征细粒度增强的鲁棒乳腺超声肿瘤分割模型，并进一步构建基于注意力机制的多区域联合分割架构。（2）高准确性的乳腺超声肿瘤良恶性分类新模型。包含基于多尺度、多核特征选择的集成分类模型；基于传统/深度多维特征融合的肿瘤分类架构和基于多视点、多约束学习的肿瘤良恶性分类模型。研究成果将有望为乳腺超声肿瘤临床良恶性辅助诊断提供新的技术手段，有效提升辅助诊断的准确性。

Ultrasound image-assisted diagnosis of breast cancer has important clinical significance in screening breast diseases, assisting the detection of tumors and improving the accuracy of diagnosis. The project intends to study the key technologies of breast ultrasound image assisted diagnosis, and study new methods of breast ultrasound image segmentation and classification based on multi-dimensional features. It mainly includes: (1) Automatic segmentation of breast tumors and related areas in ultrasound images. A robust breast ultrasound tumor segmentation model based on fine-grained enhancement of low-level features is studied, and a multi-region joint segmentation framework based on attention mechanism is further constructed. (2) New classification models of breast ultrasound tumors with high accuracy. An ensemble classification model based on multi-scale and multi-core feature selection, a tumor classification framework based on multi-dimensional fusion of traditional/deep features and a multi-view, multi-constraint learning based classification model for benign and malignant tumors are studied. The research results will hopefully provide new techniques for the clinical benign and malignant diagnosis of breast ultrasound tumors, and effectively improve the accuracy of the auxiliary diagnosis.