基于深度学习的宫颈癌淋巴结转移预测  

Prediction based on deep learning for lymph node metastasis of cervical cancer

作　　者：王佳[1] 牛俊巧[1] 李晓娟 刘焱[1] Wang Jia;Niu Junqiao;Li Xiaojuan;Liu Yan(Radioimaging Center,Xinjiang Uygur Municipal People's Hospital,Urumqi 830001,China)

机构地区：[1]新疆维吾尔自治区人民医院放射影像中心,乌鲁木齐830011

出　　处：《中国医学装备》2024年第4期71-74,79,共5页China Medical Equipment

基　　金：新疆维吾尔自治区自然科学基金(2021D01C211)。

摘　　要：目的:基于深度学习,利用T_(2)加权成像(T_(2)WI)序列的高分辨特性获得宫颈癌淋巴结的结构信息,并预测淋巴结是否转移;利用弥散加权成像(DWI)序列的功能特性,获取淋巴结区域,并预测淋巴结是否转移;综合多模态MRI数据,预测淋巴结是否转移。方法:收集2021年6月至2022年5月年新疆维吾尔自治区人民医院收治的52例宫颈癌患者的多参数MRI影像数据以及病理检查数据作为训练集,另收集2022年6月至2023年5月新疆维吾尔自治区人民医院收治的150例宫颈癌患者多参数MRI影像数据以及病理检查数据作为验证集。训练集52例宫颈癌患者均接受MRI扫描,扫描序列包括T_(2)WI和DWI序列。对训练集52例宫颈癌患者的多参数MRI影像学图像进行非均匀性校正和标准化的预处理后,通过渐进演化空洞卷积对T_(2)WI图像进行分割,在扩大感受野的同时,有效降低空洞对图像丢失的影响;通过基于注意力网络机制的深度学习模型引导网络在预测时更关注淋巴结区域,并为预测结果提供一定程度的解释性;通过多模态协同学习模型实现T_(2)WI和DWI序列在淋巴结性质预测任务之间的经验共享。采用验证集患者的图像资料对基于多模态协同学习模型的淋巴结转移预测模型进行验证。结果:验证集150例患者中良性淋巴结585枚,恶性淋巴结65枚,其良恶性淋巴结在大小(长径、短径)和边界上存在差异,差异有统计学意义(x^(2)=8.437、143.100、104.608,P<0.05)。验证集150例患者中48例患者出现淋巴结转移,基于多模态协同学习模型的淋巴结性质预测模型准确预测出46例患者出现淋巴结转移,准确预测出99例患者未发生淋巴结转移,预测准确率为96.67%。结论:渐进演化空洞卷积结合U-Net框架完成了对T_(2)WI宫颈癌影像的多目标自动分割任务,基于注意力网络机制的深度学习模型完成了动态关注淋巴结区域的任务,多模态协同学Objective:To obtain structural information on lymph nodes of cervical cancer and to predict whether lymph nodes occurred metastasis through high resolution characteristics of T_(2)-weighted images(T_(2)WI)sequences,and to obtain lymph node regions and predict whether lymph nodes occurred metastasis through using functional characteristics of DWI sequences,and to predict whether lymph occurred metastasis through synthesized multimodal magnetic resonance imaging(MRI)data.Methods:The MRI image data with multi-parameter and pathological examination data of a total of 52 patients with cervical cancer who admitted to Xinjiang Uygur Municipal People's Hospital from June 2021 to May 2022 were collected as training set,and the MRI image data with multi-parameter and pathological examination data of 150 patients with cervical cancer admitted to the same hospital from June 2022 to May 2023 were collected as verification set.All 52 patients with cervical cancer in the training set received MRI scanning,and the scanning sequence included T_(2)WI and DWI sequences.After the MRI images with multi-parameter of 52 patients with cervical cancer of the training set underwent nonuniformity correction and standardized pretreatment,the T_(2)WI were segmented by gradual evolution of hole convolution,which not only could expand the receptive field,but also effectively reduce the influence of holes on the loss of image.The network was guided to pay more attention to the lymph node region during prediction and to provide a certain extent of interpretability for the prediction results through a deep learning model based on the network mechanism of attention.Experience sharing between T_(2)WI sequences and DWI sequences in the prediction assignment of lymph node property was achieved through learning model with multimodal collaboration.The image data of patients in verification set was adopted to verify the prediction model of lymph node metastasis based on learning model with multimodal collaboration.Results:In 150 patients of the verifica

关 键 词：宫颈癌淋巴结 卷积神经网络 注意力网络机制 多模态协同学习 磁共振成像 

分 类 号：R737.33[医药卫生—肿瘤]