A Dual Radiomic and Dosiomic Filtering Technique for Locoregional Radiation Pneumonitis Prediction in Breast Cancer Patients

Abstract

Purpose: To develop a novel Explainable Dual-Omics Filtering (EDOF) model integrating dosiomic and radiomic filtering to predict locoregional radiation pneumonitis (RP) in breast cancer patients, and to explain the critical locoregional dosimetric indices and radiomic features that contribute to RP development.Materials and Methods: This retrospective study collected 72 breast cancer patients treated with radiation therapy, and a total of 28 patients developed RP (including 5 grade II cases) within 4 months post-treatment. The lung volume was first segmented from pre-treatment CT, and 3D dose distribution was also collected from the treatment plan. A 3D sliding window kernel was implemented across the (1) lung CT to capture 70 spatial-encoded image texture information, and (2) lung dose distribution to capture 36 spatial-encoded dose intensity information. As such, each voxel coordinate of the original lungs was represented as a 106-dimensional dual-omics feature vector. A novel explainable boosting machine (EBM) model was employed to establish a voxel-wise association between extracted features with locoregional RP, as identified in follow-up CT. Comparative studies against (1) radiomic filtering-only (RF) and (2) dosiomic filtering-only (DF) models were also performed. The model performance was evaluated through voxel-wise AUC, accuracy, specificity, and sensitivity with 5-fold cross-validation. The dice coefficient was additionally calculated for 5 grade II cases. The mean absolute score from EBM was also extracted to rank the feature importance. Result: The EDOF model showed highest voxel-wise RP prediction power (accuracy=0.93, sensitivity=0.93) than the RF (accuracy=0.89, sensitivity=0.01) and DF (accuracy=0.90, sensitivity=0.92) models. Specificity and AUC showed similar trends. Dice coefficient for 5 grade II patients is 0.75 in our EDOF model. Based on the EBM score, the results suggested that heterogeneous lung tissue with high locoregional dose has high risk of RP. Conclusion: The EDOF model accurately identified locoregional RP regions based on pre-treatment image and planning dose, offering a significant advancement in predictive analytics for radiation-induced complications.