3419 - Feasibility of Determining Heterogenous Risk Regions Using Dual Tree Wavelet Based Radiomics for Rectal Cancer Patients during MRgART

H. G. Nasief¹, M. Luzzara², M. P. W. Intven³, X. Chen¹, E. S. Paulson¹, and W. A. Hall¹; ¹Department of Radiation Oncology, Medical College of Wisconsin, Milwaukee, WI, ²Elekta S.p.A., Agrate Brianza (MB), Italy, ³Department of Radiotherapy, University Medical Center Utrecht, Utrecht, Netherlands

Purpose/Objective(s): Rectal cancer is one of the leading causes of cancer deaths with a low five-year survival rate if the cancer spreads to distant parts of the body. Previously we showed that wavelet delta-radiomics is a very efficient multiscale analysis method that can predict treatment outcomes for pancreatic cancer during MR-guided adaptive radiation therapy (MRgART). This work aims to investigate the feasibility of using dual tree wavelet radiomics to benchmark textures associated with heterogenous risk regions for rectal cancer patients. Materials/

Methods: Daily MR images acquired during MRgART using a 1.5 T MR Linac for 25 rectal cancer patients were analyzed retrospectively. These data were collected through an international multi-institutional registry (MOMENTUM, NCT04075305). For each MRI set, the GTV and the surrounding rectum tissue were manually contoured. The region enclosing the rectum-GTV was generated using commercially available software. Dual tree wavelet transform was used to characterize the differences between the GTV and the surrounding tissue. Radiomic textures were calculated from different decomposition levels for each ROI. Self-organized neural network map (SOM) was used to cluster these textures into corresponding groups. T-test and decision trees were used to determine features that can be used to distinguish the GTV.
Results: Overall, dual tree wavelet maps have the potential to distinguish heterogenous high-risk regions corresponding to GTV. SOM demonstrated that 8 wavelet radiomics textures can be used to benchmark heterogeneous risk regions. Four of these textures, HLL_Entropy(p-value=0.024), HLL_Cluster tendency(p-value=0.0045), HHL_ Maximum probability (p-value=0.0001), and HHL_gray level nonuniformity (p-value=0.0045), passed the t-test and demonstrated significant difference between GTV and surrounding tissue. Decision tree combing these four textures could distinguish high-risk regions with an AUC of the ROC of 0.96.
Conclusion: Using dual tree wavelet radiomics to distinguish heterogenous risk region is feasible. A larger verification study is needed to develop this into a clinical tool that can aid in treatment adaptation for better patient-specific outcomes.