基于多数据融合深度网络的甲状腺肿瘤精准诊断

Thyroid cancer is the most common thyroid malignancy, the number of patients is increasing dramatically. At present, the clinical diagnosis of thyroid cancer before operation is mainly based on ultrasound and cell puncture to determine the type of the tumor, and tissue sections examination during operation. The main problems of preoperative diagnosis are that Ultrasound can not identify the type of tumor accurately and there are many limitations in the use of cell puncture. The main problem of intraoperative diagnosis is that it takes a long time to diagnose and the results can only be obtained after operation, which affects the timeliness of treatment. This project intends to use artificial intelligence to combine ultrasound, cell and tissue images together to diagnose thyroid tumors comprehensively. By employing the integration of deep learning and traditional handcrafted feature, A multi-image source, multi-dimension, multi-resolution feature extraction, fusion and diagnosis of tumors can be realized. It can not only make accurate type diagnosis of tumors before operation, but also improve the efficiency of intraoperative pathological detection. Finally, it is of great practical significance to provide a practical intelligent analysis system for the medical thyroid auxiliary examination, to help clinicians reduce the occurrence of over-treatment and misdiagnosis, and to solve the difficult medical problems of accurate pathological diagnosis and shortage of pathologists.

甲状腺癌是最常见的甲状腺恶性肿瘤，近年来发病率快速攀升，每年以6%的速度递增。目前术前超声检查不能明确肿瘤类型，细胞穿刺在使用上有很多限制。术中组织切片诊断时间长，往往在术后才能得到结果，影响治疗的及时性。本项目拟采用深度神经网络融合超声，细胞和组织图像对甲状腺肿瘤进行精准诊断。设计多尺度网络对疑似病变区域进行检测与分割，依据临床和病理医生诊断的评判准则设计特征描述子，提出具有前端、后端和全融合三种方式的深度神经网络对多源图像、多特征进行融合，实现甲状腺肿瘤的精准诊断。通过上述方法创新，在术前、术中和术后等不同应用场景，实现融合不同甲状腺图像的高精度、细粒度诊断。项目的研究将为甲状腺诊断提供一套切实可行的智能分析系统，协助临床医生减少过度医疗及漏诊误诊的发生，解决超声、病理精确诊断难度大、相关医生短缺等难题，具有非常广泛的应用前景。

项目的研究目标是采用深度神经网络融合图像、文本等多种模态数据对甲状腺肿瘤进行精准诊断。设计多尺度网络对疑似病变区域进行检测与分割，依据临床和病理医生诊断的评判准则设计特征描述子，提出具有前端、后端和全融合三种方式的深度神经网络对多源图像、多特征进行融合，实现甲状腺肿瘤的精准诊断。. 课题组围绕上述目标展开研究，收集甲状腺肿瘤超声图像数据6000多例，根据美国放射协会提出的标准分为1-5级，并对图像中肿瘤的轮廓位置和类型都进行了标注；创新性提出了一种两阶段甲状腺肿瘤分割算法，提升边缘分割效果；探索了文本与图像信息融合的算法框架，将文字信息通过特征提取后，与图像特征进行拼接，提升甲状腺结节分割效果。. 项目总共发表高水平论文74篇，其中《IEEE Trans. on Medical Imaging》、《IEEE Trans. on Image Processing》等SCI中科院1区论文30篇；医学图像处理顶会MICCAI 2020/2022国际会议论文5篇，计算机视觉顶会CVPR 2020/2021/2022、ICCV 2021、ECCV 2020/2022以及人工智能/多媒体顶会AAAI 2021、IJCAI 2022和ACM Multimedia 2021/2022论文18篇，申请发明专利2项，完成发明专利转让1项。培养博士后出站4名，硕士研究生毕业23名。. 项目负责人沈琳琳教授在项目执行期连续三年（2020-2022）入选“Elsevier中国高被引学者”，2020-2022连续入围斯坦福大学“科学影响力全球前2%科学家榜单”，2020年1月获聘深圳市“鹏城学者”特聘教授。