Author: Laurence Edward Court, Alexandra Olivia Leone, Zhongxing Liao, Saurabh Shashikumar Nair, Joshua S. Niedzielski, Ramon Maurilio Salazar, Ting Xu 👨🔬
Affiliation: The University of Texas MD Anderson Cancer Center, Department of Radiation Physics, The University of Texas MD Anderson Cancer Center 🌍
Purpose: Radiation Pneumonitis (RP) predictive models often rely on clinical and DVH parameters, but multiomic features from CT imaging and 3D dose distributions from various regions could provide additional information. We investigate the utility of radiomic and dosiomic features from these regions to improve radiation pneumonitis toxicity models.
Methods: 329 NSCLC patients treated with IMRT, (n=190), passive-scatter proton therapy (n=92), and intensity-modulated proton therapy (n=47) were collected. Patient toxicity ≥grade 2 RP was assessed using CTCAE v5.0. Radiomic and dosiomic features were extracted from the total lung volume minus GTV as well as from a 20-mm ring subregion of normal lung tissue surrounding the GTV from the planning CT and the 3D dose volume, respectively, using pyradiomics (v3.0.1). A total of 423 features (13 Clinical, 18 dosimetric, 182 dosiomic, 210 radiomic) were extracted. Four toxicity prediction model types were created using clinical factors together with one of the following: (a) base DVH, (b) whole-lung radiomic & dosiomic (WL-RD), (c) whole-lung and ring (multi-region) radiomic & dosiomic (MR-RD), and (d) multi-region DVH + radiomic + dosiomic (MR-DVHRD). Toxicity models were created using a random forest classifier and a repeated cross-validation approach of 100 iterations, with a training/test split of 80%/20%, respectively.
Results: Model predictive performance was evaluated by area under the curve (AUC) and area under the precision-recall curve (AUPRC) which were: (a) 0.81/0.70 (base DVH), (b) 0.82/0.73 (WL-RD, p<0.05), (c) 0.83/0.75 (MR-RD, p<0.00001) and (d) 0.82/0.72 MR-DVHRD, p<0.05) respectively. P-values are the Wilcoxon sign-ranked paired difference AUC/AUPRC from the base DVH model and the respective multiomic model type.
Conclusion: All multiomic models outperformed the base model. MR-RD was the best performing model type and dosiomic features were the most prevalent in all multiomic model types. Radiomics and dosiomics in the multi-region approach provides novel insights for RP management.