基于深度学习T2WI及弥散加权成像双模态影像一体化模型自动识别及分割宫颈癌

Integrated model of T2WI and diffusion weighted imaging based on deep learning for automatic dual-mode recognition and segmentation of cervical cancer

目的基于深度学习(DL)结合Transformer网络及卷积神经网络(CNN)构建T2WI及弥散加权成像(DWI)双模态宫颈癌影像自动识别及分割一体化模型,并观察其应用价值。方法回顾性收集116例经病理确诊的宫颈癌患者,对其中58例基于盆腔轴位T2WI、80例基于盆腔轴位DWI手动勾画肿瘤ROI,之后行2D切片,标注为“肿瘤”或“非肿瘤”,共获得1166幅T2WI和1066幅DWI 2D切片。随机选取200幅T2WI(46幅肿瘤切片及154幅非肿瘤切片)和174幅DWI 2D切片(62幅肿瘤及112幅非肿瘤)为测试集,按4∶1比例将其余966幅T2WI和892幅DWI 2D切片分为训练集和验证集。以Swin Transformer网络构建宫颈癌四分类自动识别模型,结合迁移学习方法,对训练集和验证集的2个模态切片进行分类。基于nnU-Net框架开发2个通道深度分别为7层与8层的U-Net网络,构建不同模态影像宫颈癌自动分割模型;根据准确率(ACC)、精确度(Precision)、召回率(Recall)和平衡F分数(F1-score)评估模型自动识别测试集宫颈癌的效能,以戴斯相似性系数(DSC)、95%豪斯多夫距离(95%HD)及平均表面距离(MSD)评价其自动分割测试集宫颈癌的效能。结果自动识别模型识别测试集T2WI及DWI 2D切片中的宫颈癌的总体ACC、Recall、Precision及F1-score分别为86.90%、69.44%、82.42%及0.75。自动分割模型分割测试集T2WI 2D切片中的宫颈癌的DSC、95%HD及MSD均值分别为76.69%、14.85 mm及4.10 mm,分割DWI 2D切片中的宫颈癌的DSC、95%HD及MSD均值分别为84.18%、3.28 mm及0.42 mm。结论DL结合Transformer网络及CNN构建的T2WI及DWI双模态影像一体化模型能有效自动识别并分割宫颈癌病灶。

Objective To integratedly construct an automatic recognition and segmentation model of T2WI and diffusion weighted imaging(DWI)based on deep learning(DL)for cervical cancer,combined with Transformer network and convolutional neural network(CNN),and to explore its application value.Methods A total of 116 patients with pathologically diagnosed cervical cancer were retrospectively collected.Tumor ROIs were manually delineated based on pelvic axial T2WI of 58 cases and pelvic axial DWI of 80 cases.Totally 1166 T2WI and 1066 DWI 2D slices were obtained and labeled as"tumor"or"non-tumor".Then 2002D T2WI(46 tumor and 154 non-tumor)and 1742D DWI(62 tumor and 112 non-tumor)slices were randomly selected as the test set,and the remaining 966 T2WI and 892 DWI 2D slices were divided into training set or validation set at the ratio of 4∶1.Swin Transformer network was used to construct a four-classifier model for automatic recognition of cervical cancer.Combined with transfer learning,the model was used for recognition of two modal slices in training set and validation set.U-Net networks with channel depth of 7-layer and 8-layer were developed based on nnU-Net,respectively,to construct automatic segmentation model for cervical cancer in different modalities.The accuracy(ACC),precision,recall(Recall)and balanced F-score(F1-score)were applied to evaluate the model performances for automatic detection of cervical cancer in the test set.Dice similarity coefficient(DSC),95%Hausdorff distance(95%HD)and the mean surface distance(MSD)were used to assess the model performances for automatic segmentation of cervical cancer in test set.Results The overall ACC,Recall,Precision and F1-score of automatic recognition model to identify cervical cancer for T2WI and DWI 2D slices in test set were 86.90%,69.44%,82.42%and 0.75,respectively.The mean values of DSC,95%HD and MSD of automatic segmentation model to delineate cervical cancer area for T2WI 2D slices in test set were 76.69%,14.85 mm and 4.10 mm,respectively,while for DWI 2D slices were 84.18%,3.28 mm and 0.42 mm,respectively.Conclusion The DL-based integrated model of T2WI and DWI images combined with Transformer network and CNN could effectively identify and segment cervical cancer lesions.