Negligible effect of brain MRI data preprocessing for tumor segmentation

Ekaterina Kondrateva, Polina Druzhinina, Alexandra Dalechina, Svetlana Zolotova, Andrey Golanov, Boris Shirokikh, Mikhail Belyaev, Anvar Kurmukov

Abstract: Highlights•Deep CNN and visual transformers do not benefit from image bias-field correction, denoising or harmonization.•Image resizing or voxel resampling produce comparable results to atlas registration.•Skull stripping improves Dice scores but has limited effect on tumor volumetric measurements.•Preprocessing MRI data does not aid transfer learning while fine-tuning on other datasets.•Fine-tuning on other data or longer training on original data could yield similar results on small heterogeneous samples.

External IDs:dblp:journals/bspc/KondratevaDDZGSBK24