利用双注意力CycleGAN从超声合成数字乳腺断层扫描病灶研究  

Synthesized lesion of digital breast tomosynthesis from ultrasound using dual attention CycleGAN

作　　者：马健晖 唐钒 梁宇棋 李晏宁 刘书朋 齐宏亮 Ma Jianhui;Tang Fan;Liang Yuqi;Li Yanning;Liu Shupeng;Qi Hongliang(Department of Radiotherapy,Southern Hospital,Southern Medical University,Guangzhou 510515,China;Department of Medical Engineering,Southern Hospital,Southern Medical University,Guangzhou 510515,China)

机构地区：[1]南方医科大学南方医院放疗科,广州510515 [2]南方医科大学南方医院医学工程科,广州510515

出　　处：《现代仪器与医疗》2023年第4期63-69,79,共8页Modern Instruments & Medical Treatment

基　　金：国家自然科学基金青年项目(82202960);广东省医学会一般项目(2022-GDMAYB-04)。

摘　　要：目的为了提高乳腺癌临床诊断准确性的同时,降低数字乳腺断层扫描(Digital Breast Tomosynthesis,DBT)带来的辐射剂量风险,提出一种从超声等非放射性医学成像中合成DBT图像的方法。方法由于DBT和超声在组织形态学上存在巨大差异,本文首先研究了DBT和超声中乳腺病灶的有效代表,然后提出了一种双注意力CycleGAN框架,将空间和通道注意机制与CycleGAN相结合,用于合成DBT病灶。结果定性和定量结果表明,本文方法能够合成接近真实图像质量的DBT图像,良恶性诊断的精确度、灵敏度和特异性分别为0.83、0.71和0.90。结论本文提出的方法具有从超声合成DBT图像的优异性能潜力。Objective To improve the accuracy of clinical diagnosis of breast cancer and reduce the risk of radiation dose irradiated by digital breast tomosynthesis(DBT),this study proposes a method to synthesize DBT images from nonradioactive medical imaging such as ultrasound.Methods Since DBT and ultrasound are vastly different from each other in histomorphology,this paper firstly investigates the effective representative of breast lesion in DBT and ultrasound,then proposes a dual attention CycleGAN framework integrating spatial and channel attention mechanism with CycleGAN for intractable DBT lesion synthesis.Results The qualitative and quantitative results demonstrate that this method can synthesize DBT images close to the real image quality.The accuracy,sensitivity,and specificity of benign and malignant diagnosis are 0.83,0.71,and 0.90,respectively.Conclusion The proposed method has potential to achieve superior performance of DBT synthesis from ultrasound.

关 键 词：数字乳腺断层扫描 超声 双注意力机制 生成对抗网络 

分 类 号：TH774[机械工程—仪器科学与技术] R81[机械工程—精密仪器及机械]