Abstract:
Whether applied for clinical research or patient health risk assessment, our aim is to implement a brain tumor classification and segmentation approach, with a focus on extracting tumor shape and texture features and investigating potential associations with genomic subtypes. By using a combination of UNET with ResNeXt50 backbone architecture, we investigate the improvement of model performance on a basis of hyperparameter alteration, as well as determining statistically significant associations within lower grade gliomas. We achieved a Mean Dice accuracy of 95% with the UNET ResNeXt50 model in tumor segmentation and in terms of extracting radiomic features. Our strongest shape feature associations across all three types of tumors resulted between Bounding Ellipsoid Volume Ratio and RNASeqCluster (p<0.008), RPPACluster
(p<0.002); Convexity Defects and CNCluster (p<0.001), COCCluster (p<0.04); Correlation and RPPACluster (p<0.03); Homogeneity and RNASeqCluster (p<0.001), MethylationCluster (p<0.0003), OncosignCluster (p<0.002); Energy and RPPACluster, MethylationCluster (p<0.001). Our ROC AUC scores, pointed out the best discriminative abilities found in BEVR, Equivalent Diameter, Contrast for CNCuster C3 and RPPACluster R4, as well as Extent and Convexity Defects for Methylation Cluster M1.
