深度卷积网络模型可自动识别与分割胰腺及其肿瘤:基于3D V-Net

A deep convolutional network model based on 3D V-Net can automatically recognize and segment pancreas and its tumors

目的探讨基于V-Net的深度卷积神经网络模型在胰腺及其肿瘤自动识别和分割任务中的有效性和可行性。方法回顾性分析2012年5月~2019年11月于上海交通大学医学院附属第一人民医院就诊且经病理证实为胰腺癌的186例患者的增强CT影像资料,经过筛选,共纳入108例胰腺癌病例,随机搜集同期37例正常胰腺病例用于对照,最终共纳入145例数据,构成本研究的数据集。采用五折交叉验证方法,在动脉期CT图像上进行人工标注感兴趣区域(ROI),包括胰腺头颈部、体尾部和肿瘤,通过计算敏感度、特异度、F1分数等指标评估模型对胰腺肿瘤的识别能力,并进行Kappa一致性验证。采用Dice系数定量评估模型的分割能力,并获取可视化结果进一步评估。结果基于V-Net的模型识别胰腺肿瘤的敏感度为0.852、特异度为1.000、阳性预测值为1.000、阴性预测值为0.698,F1分数高达0.920。一致性验证显示,Kappa系数为0.746(P<0.05)。在分割任务中,胰腺肿瘤、胰腺体尾部和胰腺头颈部的Dice系数分别为0.722±0.290、0.602±0.175、0.567±0.200。结论本研究构建基于VNet的深度卷积网络模型,有效完成了胰腺及其肿瘤自动识别与分割,验证了该方法的有效性和可行性,为探索胰腺肿瘤领域人工智能应用提供有力支持。

Objective To explore the effectiveness and feasibility of the deep convolutional neural network model,based on VNet,for automatic recognition and segmentation of the pancreas and its tumors.Methods A retrospective analysis was conducted on the enhanced CT imaging data of 186 patients with pathologically confirmed pancreatic cancer who visited First People's Hospital Affiliated to Shanghai Jiaotong University Medical College from May 2012 to November 2019.After screening,a total of 108 cases of pancreatic cancer were included,and 37 cases of normal pancreas during the same period were randomly collected for comparison,resulting in a final dataset of 145 cases for this study.This paper employed a five-fold cross-validation method and manually annotated regions of interest on arterial phase CT images,including the pancreatic head and neck,body and tail,and tumors.The model's ability to identify pancreatic tumors was evaluated by calculating metrics such as sensitivity,specificity,F1 score,and Kappa consistency verification was performed.Dice coefficient was used to quantitatively assess the model's segmentation capability,and visual results were obtained for further evaluation.Results The V-Net based model for identifying pancreatic tumors has a sensitivity of 0.852,a specificity of 1.000,a positive predictive value of 1.000,a negative predictive value of 0.698,and an F1 score as high as 0.920.The consistency verification shows that the Kappa coefficient is 0.746(P<0.05).In the segmentation task,the mean Dice for pancreatic tumors,pancreatic body and tail,pancreatic head and neck were 0.722±0.290,0.602±0.175,0.567±0.200,respectively.Conclusion We constructed a deep convolutional network model based on V-Net,which successfully achieved automatic identification and segmentation of the pancreas and tumors.Our findings demonstrated the effectiveness and feasibility of this approach,offering robust support for the exploration of artificial intelligence applications in the field of pancreatic tumor research.