Pseudo-siamese network combined with dosimetric and clinical factors, radiomics features, CT images and 3D dose distribution for the prediction of radiation pneumonitis: A feasibility study. (January 2023)

• Record Type: Journal Article
• Authors: Feng, Bin; Zhou, Wei; Yang, Xin; Luo, Huanli; Zhang, Xin; Yang, Dingyi; Tao, Dan; Wu, Yongzhong; Jin, Fu
• Is Part Of: Clinical and translational radiation oncology. Volume 38(2023)
• Page Start: 188
• Page End: 194
• Publication Date: 2023-01
• Subjects: Radiation pneumonitis -- Pseudo-siamese network -- Radiomics -- Radiotherapy -- Lung cancer; Cancer -- Radiotherapy -- Periodicals; Oncology -- Periodicals; Cancer -- Radiotherapy; Oncology; Radiation Oncology; Neoplasms -- radiotherapy; Translational Medical Research; Periodicals; Electronic journals; Periodicals; 616.9940642
• DOI: 10.1016/j.ctro.2022.11.011
• Languages: English
• ISSNs: 2405-6308

Abstract

Highlights: We firstly proposed a novel method that combined dosimetric and clinical factors, radiomics feature, 3D CT images and dose distribution for predicting RP. A pseudo-siamese network were developed for combination of 1D and 3D information. The used deep learning frame showed a good accuracy (81.8%) for the prediction of RP. Abstract: Purpose: Radiation pneumonitis (RP)(grade ≥ 2) can have a considerable impact on patient quality-of life. In previous studies, the traditional method commonly used radiomics and clinical factors for RP prediction. This study aims to develop and evaluate a novel pseudo-siamese network (PSN) to assist radiologists predict RP before radiotherapy based on combination of dosimetric and clinical factors, radiomics features, CT (computed tomography) images, and dose distribution (hybrid model). Method: One hundred and ten patients with lung cancer (19 RP ≥ 2) who received radiotherapy between 2016 and 2020 were retrospectively enrolled in this study. Dosimetric factors were calculated from DVH (dose-volume histogram), such as lung mean dose, lung V5, and prescription dose. Clinical characteristics were recorded, such as age, sex, smoking status, TN stage, and overall stage. A total of 1419 radiomics features were extracted. Cluster analysis was used for detecting radiomics features that associated with RP. Patients were randomly split into a training set (90 %, 85 non-RP, and 14 RP) and a validation set (10 %, 6 non-RP, and 5 RP). A PSN architecture was designed for combining 1D (dosimetric and clinical factors, radiomics) and 3D (CT images, 3D dose distribution) features. 5-fold cross-validation procedure for estimating the skill of the model on new data. Results: For cluster analysis, totally of 106 radiomics features with high correlation were selected. The accuracy was 0.727, 0.636, 0.545, and 0.727 for input dosimetric and clinical factors, dose distribution, CT images, and radiomics features, respectively. The accuracy of hybrid model was 0.818. The sensitivity of hybrid model was 0.800 (95 % confidence interval (CI) [0.299, 0.989]), and specificity was 0.833(95 % CI [0.364, 0.991]). The areas under the receiver operating characteristic curves (AUCs) result in 5-fold cross-validation was 0.77–0.90(mean AUC ± std was 0.85 ± 0.05). Conclusion: This study firstly propose method that the combination of high dimensional and low dimensional features for RP prediction. The results confirm the feasibility of multi-dimensional features predict RP.