Fig. 1

Proposed radiomics pipeline. a Radiomics feature generation pipeline. I: Deep learning based segmentation, II: Connected component labeling removing isolated mask elements, III: Manual consistency check of segmentation masks, IV: Mask erosion to remove boundary elements, V: Texture feature extraction of each muscle group, VI: Shape feature extraction of each muscle group, VII: fat fraction calculation based on water and fat contrast per muscle mask. b Radiomics feature selection pipeline with five-fold cross-valdiation for steps II-V. I: Cluster representative identification based on correlation values, II: Discarding features with low mean left and right muscle correlation values, III: Variance inflation factor calculation and iterative feature removal, IV: Heuristic Boruta selection to estimate feature importance with respect to surrogate targets body mass index (BMI), age and fat fraction (FF), V: Permutation importance calculation across folds for identification of ranking of selected features. Numbers on the arrows indicate the remaining amount of texture features after each selection step