Automatic pancreatic cancer GTV segmentation based on deep learning
10.3969/j.issn.1005-202X.2025.07.012
- VernacularTitle:基于深度学习的胰腺肿瘤靶区自动分割
- Author:
Chaoshuang CHEN
1
;
Yangsen CAO
;
Xiaofei ZHU
;
Fubin ZENG
;
Lei GU
;
Lingong JIANG
;
Huojun ZHANG
Author Information
1. 中国人民解放军92493部队医院核医学肿瘤科,辽宁 葫芦岛 125000
- Publication Type:Journal Article
- Keywords:
pancreatic cancer;
gross tumor volume;
segmentation;
deep learning;
convolutional neural network
- From:
Chinese Journal of Medical Physics
2025;42(7):923-928
- CountryChina
- Language:Chinese
-
Abstract:
Objective To investigate the feasibility and accuracy of convolutional neural networks for automatically delineating the pancreatic cancer gross target volume(GTV)in pancreatic enhanced CT.Methods The localizable enhanced CT images of 114 patients with pancreatic cancer were retrospectively selected,in which the GTV was manually delineated using AccuContour.The imaging data were then import to AccuLearning and randomly divided as the training set,validation set and test set at a ratio of 8:1:1.Flex and Segresnet were used to train the automatic segmentation model,with each network structure trained continuously 3 times using fixed training parameters.The model was evaluated in terms of Dice similarity coefficient(DSC),95%Hausdorff distance(HD95),average symmetric surface distance(ASSD)and relative volume difference(RVD).Results In the model training phase,Flex-3 test results in Flex group were the worst,with a minimum DSC of 0.14%and an average DSC of 56.30%,while Flex-1 performed well,achieving a minimum DSC of 47.90%and an average DSC of 67.35%.Meanwhile,Segresnet-2 in Segresnet group had the worst test results,with a minimum DSC of 0.00%and an average DSC of 42.46%,while Segresnet-3 test results were better,with a minimum DSC of 42.65%and an average DSC of 63.28%.In the fixed testing phase,the best results among all were as follows:average DSC and RVD values of 63.88%and 29.41%in Segresnet-3 group,average ASSD value of 4.43 mm in Segresnet-2 group,and average HD95 value of 12.87 mm in Segresnet-1 group.Conclusion Both Flex and Segresnet architectures of convolutional neural network can be used for the automatic pancreatic tumor GTV segmentation training,with Segresnet performing better in comprehensive evaluation.
