基于组织定征和超声组学构建甲状腺癌淋巴结转移风险预测模型的研究

The incidence of thyroid cancer is increasing year by year, and cervical lymph node metastasis is the most important risk factor affecting recurrence and distant metastasis, and lymphatic metastasis decides the treatment plan directly. At present, there is no reliable diagnostic method to evaluate the risk of cervical lymph node metastasis in thyroid carcinoma. Related studies have found that the risk of lymph node metastasis in thyroid cancer is related to tumor shape and position, but the lack of a quantitative model to consider multiple risk factors, so the prediction accuracy rate is low, the contribution for clinical treatment decision is limited. This project creatively uses the ultrasonic issue characterization and radiomics analysis method, combined with the big data analysis technology, from 1000 cases and tens of thousands of ultrasound images and ultrasonic raw data mining the correlation characteristic, uses the network meta analysis method from the clinical data analysis summarizes the lymph node metastasis related risk factor. Using the method of deep learning combined with probabilistic graph model, this research constructs a quantitative predictive model of lymph node metastasis risk of thyroid carcinoma, establishes a technical basis for solving the problem of quantitative prediction of lymph node metastasis in thyroid carcinoma, and promotes the improvement of diagnosis and treatment of clinical thyroid cancer.

甲状腺癌发病率逐年增高，颈部淋巴结转移是影响患者复发和远处转移的最重要危险因素，淋巴转移情况直接决定治疗方案的制定。而现阶段仍没有一个可靠的诊断方法可以有效评估甲状腺癌颈部淋巴结转移风险程度。相关研究已经发现甲状腺癌淋巴结转移风险与肿瘤形态、位置等因素相关，但缺乏一种定量化的模型将多种风险因素综合考虑，因此预测准确率低，对临床治疗方案的确定帮助有限。本项目创新性利用组织定征算法和超声组学技术，结合大数据分析技术，从上千例甲状腺恶性结节的近万幅超声影像以及超声原始数据中挖掘相关特征，用网络Meta分析方法从临床数据中分析总结出与淋巴结转移相关的风险因素，用深度学习结合概率图模型方法最终将多种风险因素和超声特征综合，构建甲状腺癌淋巴结转移风险量化预测模型，为解决甲状腺癌淋巴结转移风险定量预测的难题奠定技术基础，推动临床甲状腺癌诊断和治疗方法的完善。

甲状腺癌发病率逐年增高，颈部淋巴转移情况直接决定治疗方案的制定。而现阶段仍没有一个可靠的诊断方法可以有效评估甲状腺癌颈部淋巴结转移风险程度。本项目创新性利用超声组学技术，结合大数据分析技术，从上千例甲状腺恶性结节的近万幅超声影像以及超声原始数据中挖掘相关特征，并结合深度学习方法构建甲状腺癌淋巴结转移风险量化预测模型，为解决甲状腺癌淋巴结转移风险定量预测的难题奠定技术基础。项目重点研究了占甲状腺癌80%的甲状腺乳头状癌（Papillary Thyroid Cancer，PTC）的颈部淋巴结转移情况，建立了基于深度卷积神经网络的预测模型。用来自3,794例PTC的17,356幅超声图像进行训练，在三家医院：浙江大学医学院第二附属医院、杭州市第一人民医院、上海市第十人民医院的多中心数据中进行测试，测试分为单发病灶（339例）和多发病灶（364例）两个队列。研究得出的结论和关键数据如下：对于单发PTC深度学习的预测敏感性87.6% (297/339; 95% CI: 84.5%–89.1%)，特异性为84.1% (285/339; 95% CI: 81.2%–85.9%), 曲线下面积AUC值为0.911 (p < 0.001)。对于多发PTC深度学习的预测敏感性为 85.2% (310/364; 95% CI: 81.9%–87.0%)，特异性为82.4% (300/364; 95% CI: 80.1%–83.7%), 曲线下面积AUC值为0.893 (p < 0.001)。研究取得的重要结论为：通过结合超声图像语义特征提取，建立基于深度学习网络的模型可以较为准确的预测甲状腺乳头状癌的颈部淋巴结转移情况，对于甲状腺癌的临床检查和后续治疗具有重要指导作用。