Swin-UNETR pretrain roi size

[benecia2sj]

Thank you for sharing the pretraining code for Swin-UNETR. I’m planning to pretrain Swin-UNETR using MRI scans of various body parts (e.g., brain, abdomen) and would appreciate your insights on the impact of roi size (x, y, z) during pretraining.

The default roi size is 96, and I’m curious about how adjusting this size—either increasing or decreasing—affects the performance of downstream tasks like segmentation. My understanding is that the roi needs to be large enough to encompass a significant portion of the organs to be effective.... Was the default value of 96 determined to be optimal based on your experiments, or would you recommend adjustments depending on the specific application? Thank you.