3087 - Developing an MRI Radiomic Fingerprint for Prognostic Intra-Tumor Heterogeneity for Rectal Cancer Patients Treated on the 1.5T MR-Linac

Purpose/Objective(s): The response of rectal cancer to radiotherapy is variable. Previously we showed that dual tree wavelet radiomic maps have the potential to distinguish heterogeneous risk regions capable of differentiating the GTV from the surrounding rectal tissue during MR guided adaptive radiation therapy (MRgART). This work aims to expand the study on a larger cohort of patients and to investigate the feasibility of developing a delta radiomic fingerprint (DRF) for rectal cancer patients treated on the 1.5T MR-linac to highlight its potential role in response assessment during therapy.

Materials/Methods: Daily T2 weighted MR images acquired during MRgART using 1.5 T MR Linac for 65 rectal cancer patients treated in 5 fractions were analyzed retrospectively. These data were collected through an international multi-institutional registry (MOMENTUM, NCT04075305). Patients were divided into two groups based on the lymph node involvement. For each MRI set, for primary tumor a GTV and a ring-shaped region around the GTV were used. For lymph nodes, lymph node GTV (GTV_LN) and a ring expansion around it was used. A Boolean operation was used to generate ring minus GTV structures (‘RING’_ ‘GTV’, ‘RING’_ ‘GTV_LN’) for the analysis. Contour QA was performed to exclude non tissue voxels, i.e. voxels with stool or mucus etc. prior to feature extraction. Radiomic features were extracted from each ROI from longitudinal images. DRFs were calculated for each ROI as the relative net change of the feature value compared to the first fraction. T-tests, linear mixed effect model, and logistic regression models were used to determine features with significant differences between the GTV and the ring. Significant DRFs fingerprint were used to build a naïve Bayesian classifier. Performance was judged using a leave one out cross validated AUC (CV-AUC) of the ROC curve.

Results: Five DRFs passed the t-test and linear mixed effect model with p-value <0.05, demonstrating significant differences between the GTV, GTV_LN and the corresponding ring shells. Logistic regression model demonstrated increased significance differences for Skewness, Mean and Kurtosis DRFs by mid treatment. An optimizable naïve Bayesian model combining 3-DRFs (Skewness, Kurtosis, Mean) distinguished the GTV from the ring shell with a CV-AUC of 0.84 and GTV_LN from the corresponding ring shell with CV-AUC of 0.92.

Conclusion: Developing a delta radiomic fingerprint of rectal tumors during MRgART on the 1.5T MR-linac is feasible. A larger verification study is needed to incorporate these delta-radiomic features with the dual tree wavelet maps and clinical biomarkers to develop this into a clinical tool that can aid in determining treatment response.