search

uni-medical / SAM-Med3D

SAM-Med3D: An Efficient General-purpose Promptable Segmentation Model for 3D Volumetric Medical Image
Apache License 2.0
531 stars 71 forks source link

Possible bias in training #70

Closed Lucashsmello closed 2 months ago

Lucashsmello commented 5 months ago

Hi! Thanks a lot for the code!

Does using torchio.CropOrPad(mask_name='label') in the training phase not contribute to a bias? Because of always centering the image on the label.

blueyo0 commented 2 months ago

Indeed, your observation is correct. This version is designed with an excessive focus on the center. While this approach proves detrimental for targets exceeding 128 pixels, it has negligible impact on small to medium-sized targets.

If your focus is large targets, further fine-tuning is needed.