基于ABVS图像的乳腺病变检出与分类及参数测量

Recently the automated breast volume scanner (ABVS) has been applied in clinical and drawn attention in breast disease detection and diagnosis applications. Relevant researches in objective analysis and parameter measurement of breast cancer based on ABVS images are in the start-up stage at home and abroad. There exist some difficulties and problems. Noise produced by nonuniform coupling agent and fatty tissues both display low echo property, leading to a false positive region show tumor characteristics in one section. The edge description is normally employed to differentiate benign tumor from malignant tumor in clinical. However, using hand-crafted features to quantitatively express edge description is hard. The tumor boundary is obscured due to partial volume effect and gradual change of the biology tissue, which increases the tumor segmentation difficulty. Parameter measurement and three-dimensional reconstruction can aid clinician to design preoperative surgery plan. However, the relevant research is rare. Against this background, the study on the detection, classification and parameter measurement for breast tumor will be conducted in this project using the ABVS images of Chinese women. The research contents include: suspicious tumor region detection, suspicious tumor region classification, tumor edge segmentation, parameter measurement and three-dimensional reconstruction. This project refers to the theories and methods of information science and medical science. It will lay a foundation for objective diagnosis of breast cancer and reasonable design of surgery.

近年来全自动乳腺容积扫描(ABVS)开始应用于临床，其在乳腺病变检出以及诊断中的应用逐渐受到关注。基于ABVS图像的乳腺癌客观分析以及参数测量的相关研究在国内外尚属起步阶段，存在以下难点和问题：耦合剂不均匀产生的噪声与脂肪等组织均呈现低回声特性，导致某一切面下呈现肿瘤特征的区域可能是假阳性区域；医学上一般采用边缘描述区分肿瘤良恶性，但采用人工定义特征对边缘描述进行量化表示较为困难；部分容积效应以及生物组织平缓变化的特点导致肿瘤边界不清晰，增加了肿瘤边缘分割的难度；缺乏辅助医生设计术前手术方案的参数测量以及三维重建方法的相关研究。在此背景下，本课题以国内ABVS图像为研究对象，探究乳腺肿瘤检出与分类以及参数测量方法。研究内容包括：疑似肿瘤区域检测，疑似肿瘤区域分类，肿瘤边缘分割，参数测量与三维重建。本课题结合信息学科与医学学科的理论和方法，将为乳腺癌的客观诊断与手术的合理设计奠定基础。

近年来全自动乳腺容积扫描(ABVS)开始应用于临床，其在乳腺病变检出以及诊断中的应用逐渐受到关注。基于ABVS图像的乳腺癌客观分析相关研究在国内外尚属起步阶段，存在以下难点和问题：耦合剂不均匀产生的噪声与脂肪等组织均呈现低回声特性，导致某一切面下呈现肿瘤特征的区域可能是假阳性区域；医学上一般采用边缘描述区分肿瘤良恶性，但采用人工定义特征对边缘描述进行量化表示较为困难；部分容积效应以及生物组织平缓变化的特点导致肿瘤边界不清晰，增加了肿瘤边缘分割的难度。在此背景下，本课题以国内ABVS图像为研究对象，主要研究了ABVS图像中肿瘤区域检出、边缘分割以及分类方法。基于密集连接的特征融合策略与多任务学习的目标检测模型，其肿块检测敏感度为0.90，每个切片的平均假阳性个数（FPs/S）为0.16。基于神经网络架构搜索的自动ABVS肿块分割方法，其在三维区域内肿瘤分割的Dice为0.78。基于分割模型与检测模型融合的端到端肿块分割方法，改善了三维肿块分割中目标定位难的问题，在全图范围内三维肿块分割的Dice为0.65。基于时序集成和不确定性约束的2维半监督分割模型，较只用少量有标签数据的全监督分割方法有显著提升，在仅使用300幅有标签数据时，肿瘤分割的Dice为0.73。基于特征共享与特征优化的肿块分割与良恶性分类联合训练方法，提升了感兴趣区域(Region of Interest, ROI)内三维肿块分割与良恶性分类性能，肿块分类方法的AUC为0.79。本课题结合信息学科与医学学科的理论和方法，将为乳腺癌的客观诊断与手术的合理设计奠定了基础。