Early Prediction of Radiation Treatment Response via Longitudinal Analysis of CBCT Radiomic Features for Locally Advanced Rectal Cancer.

Abstract

Patients respond to the same radiation treatment course differently due to inter- and intra- patient variability in radiosensitivity. Despite widespread use of AI/ML in radiation oncology, there is a lack of monitoring strategies used during treatment courses to evaluate early predictors of treatment response in a systematic fashion. This work advances a straightforward, yet effective, method for the early detection of treatment response through systematically analyzing daily CBCT radiomic features. The goal is to aid clinicians in assessing the treatment efficacy routinely with a view towards optimizing personalized treatment. We included a cohort of 30 patients diagnosed with locally advanced rectal cancer who underwent neo-adjuvant fractionated radiation treatment (RT) with a prescription dose of 50.4 Gy (28 fractions), followed by total mesorectal excision surgery after completion of ChemoRT. Daily IGRT imaging was acquired prior to each fraction resulting in a total of 840 CBCTs. Patients were divided into responder (14 patients) and non-responder (16 patients) groups based on post-RT pathological response. Mutual information algorithms were utilized to rigorously register daily CBCT images to the planning CT, and longitudinal radiomic features of the target were extracted from the daily CBCTs during the entire treatment course. All longitudinal features for a given patient were standardized with Z-Score normalization, followed by linear fitting using the least square method, resulting in radiomic feature trends (RFT) represented by slope values. Statistical significance was established via a two-sample U test and P-value with a threshold of 0.05. Logistic regression was performed to eliminate RFT with accuracy rates lower than 0.5. The final trending model was developed using random forest. For each patient at fraction N, our investigation involved independent 27 group experiments, where each experiment considered image group from fraction #1 to N, to confirm the effectiveness and stability of the model. The proposed RFT demonstrated a high level of precision and consistency for post-RT response based on longitudinal CBCT images for LARC patients. The trending model yielded an accuracy of 0.9556, 95% CI (0.94, 0.972) when each daily image was considered, the prediction consistency was 0.964. Given the first 14 experiments (considering group images of fraction #1-15), the prediction accuracy was 0.9357, 95% CI (0.915, 0.956) and the prediction consistency was 0.952. A strategy for monitoring and early prediction of LARC patients' radioresponse was evaluated via longitudinal CBCT assessment. Our trending models demonstrate a significant difference between the responder vs non-responder groups as early as the 15th fraction. Our strategy achieved superior accuracy and consistency to predict post-RT response of LARC patients.