3D Res2Net deep learning model for predicting volume doubling time of solid pulmonary nodule
10.13929/j.issn.1003-3289.2024.10.012
- VernacularTitle:3D Res2Net深度学习模型预测肺实性结节体积倍增时间
- Author:
Jing HAN
1
;
Lexing ZHANG
;
Linyang HE
;
Changfeng FENG
;
Yuzhen XI
;
Zhongxiang DING
;
Yangyang XU
;
Qijun SHEN
Author Information
1. 浙江康静医院放射科,浙江 杭州 310064
- Keywords:
lung neoplasms;
tomography,X-ray computed;
deep learning
- From:
Chinese Journal of Medical Imaging Technology
2024;40(10):1514-1518
- CountryChina
- Language:Chinese
-
Abstract:
Objective To observe the value of 3D Res2Net deep learning model for predicting volume doubling time(VDT)of solid pulmonary nodule.Methods Chest CT data of 734 patients with solid pulmonary nodules were retrospectively analyzed.The patients were divided into progressive group(n=218)and non-progressive group(n=516)according to whether lung nodule volume increased by ≥25％during follow-up or not,also assigned into training set(n=515)and validation set(n=219)at a ratio of 7∶3.Then a clinical model was constructed based on clinical factors being significantly different between groups,CT features model was constructed based on features of nodules on 2D CT images using convolutional neural network,and 3D Res2Net model was constructed based on Res2Net network using 3D CT images as input.Receiver operating characteristic curve was drawn,and the area under the curve(AUC)was calculated.Taken actual VDT as gold standard,the efficacy of the above models for predicting solid pulmonary nodule'VDT≤400 days were evaluated.Results No significant difference of predicting efficacy for solid pulmonary nodule'VDT≤400 days was found among clinical model,CT feature model and 3D Res2Net model,the AUC of which was 0.689,0.698 and 0.734 in training set,0.692,0.714 and 0.721 in validation set,respectively.3D Res2Net model needed 5-7 s to predict VDT of solid pulmonary nodules,with an average time of(5.92±1.08)s.Conclusion 3D Res2Net model could be used to predict VDT of solid pulmonary nodules,which might obviously reduce manual interpreting time.
