Objective:To develop and validate a combined model integrating radiomics, dosiomics, and clinical parameters based on CT simulation and dosimetric images in order to predict the occurrence of radiation pneumonitis (RP) in patients with non-small cell lung cancer (NSCLC).Methods:A retrospective study was conducted on the clinic data of 143 NSCLC patients who received radiotherapy at the Affiliated Hospital of Jining Medical University from January 2016 to December 2022. Patients were randomly stratified into a training group ( n = 100) and an internal validation group ( n = 43) at a 7∶3 ratio. Moreover, clinic data were collected from 34 NSCLC patients who received radiotherapy at the Jining Cancer Hospital between January 2019 and December 2022 as an external validation group. All three groups (the training group, internal validation, and external validation groups) were further categorized into two groups based on the RP severity (i.e., RP ≥ grade 2 and RP < grade 2). Their radiotherapy dose, CT simulation, and 3D dose distribution images were collected. Then, the total lung minus planning target volume (TL-PTV) was defined as the region of interest (ROI) for radiomics and dosiomic feature extraction, followed by feature dimensionality reduction. Consequently, key features associated with RP were determined. Four predictive models were developed using machine learning approaches (especially multilayer perceptron, MLP): a clinical model (CM), a radiomics model (RM), a dosiomics model (DM), and a radiomics and dosiomics nomogram (RDN), with a nomogram subsequently constructed. Ultimately, the performance and clinical feasibility of these models were assessed using receiver operating characteristic (ROC), area under the curve (AUC), and decision curve analysis (DCA). Results:A total of 1 834 radiomic features and 1 834 dosiomic features were extracted. Using the occurrence of RP ≥ grade 2 as the marker variable, 14 radiomic features, 15 dosiomic features, and three clinical features were selected from the training group to construct the prediction models (CM, RM, DM, and RDN). The performance and generalizability of these models were subsequently validated in both the internal validation and external validation groups. Specifically, the RDN exhibited AUCs of 0.915 (95% CI: 0.852-0.978), 0.879 (95% CI: 0.777-0.982), and 0.838 (95% CI: 0.701-0.975) in the three groups, respectively. A nomogram was established for RDN by integrating the radiomics score (R-score), dosiomics score (D-score), mean lung dose (MLD), V20, and V30. This nomogram allowed for individualized risk estimation of RP and facilitated personalized radiotherapy planning. Conclusions:The RDN model that is developed based on CT simulation and 3D dose distribution images and integrates radiomics, dosiomics, and clinical features can effectively predict the RP risk of NSCLC patients. The integration of multidimensional data contributes to the formation of the optimal predictive model, offering guidance for clinicians.

Objective:To develop and validate a combined model integrating radiomics, dosiomics, and clinical parameters based on CT simulation and dosimetric images in order to predict the occurrence of radiation pneumonitis (RP) in patients with non-small cell lung cancer (NSCLC).Methods:A retrospective study was conducted on the clinic data of 143 NSCLC patients who received radiotherapy at the Affiliated Hospital of Jining Medical University from January 2016 to December 2022. Patients were randomly stratified into a training group ( n = 100) and an internal validation group ( n = 43) at a 7∶3 ratio. Moreover, clinic data were collected from 34 NSCLC patients who received radiotherapy at the Jining Cancer Hospital between January 2019 and December 2022 as an external validation group. All three groups (the training group, internal validation, and external validation groups) were further categorized into two groups based on the RP severity (i.e., RP ≥ grade 2 and RP < grade 2). Their radiotherapy dose, CT simulation, and 3D dose distribution images were collected. Then, the total lung minus planning target volume (TL-PTV) was defined as the region of interest (ROI) for radiomics and dosiomic feature extraction, followed by feature dimensionality reduction. Consequently, key features associated with RP were determined. Four predictive models were developed using machine learning approaches (especially multilayer perceptron, MLP): a clinical model (CM), a radiomics model (RM), a dosiomics model (DM), and a radiomics and dosiomics nomogram (RDN), with a nomogram subsequently constructed. Ultimately, the performance and clinical feasibility of these models were assessed using receiver operating characteristic (ROC), area under the curve (AUC), and decision curve analysis (DCA). Results:A total of 1 834 radiomic features and 1 834 dosiomic features were extracted. Using the occurrence of RP ≥ grade 2 as the marker variable, 14 radiomic features, 15 dosiomic features, and three clinical features were selected from the training group to construct the prediction models (CM, RM, DM, and RDN). The performance and generalizability of these models were subsequently validated in both the internal validation and external validation groups. Specifically, the RDN exhibited AUCs of 0.915 (95% CI: 0.852-0.978), 0.879 (95% CI: 0.777-0.982), and 0.838 (95% CI: 0.701-0.975) in the three groups, respectively. A nomogram was established for RDN by integrating the radiomics score (R-score), dosiomics score (D-score), mean lung dose (MLD), V20, and V30. This nomogram allowed for individualized risk estimation of RP and facilitated personalized radiotherapy planning. Conclusions:The RDN model that is developed based on CT simulation and 3D dose distribution images and integrates radiomics, dosiomics, and clinical features can effectively predict the RP risk of NSCLC patients. The integration of multidimensional data contributes to the formation of the optimal predictive model, offering guidance for clinicians.