HU-based foreground masking, which trains masked image models only on subvolumes with mean intensity above 0.1, improves downstream 3D medical segmentation Dice by 0.3 to 2.9 points over random masking baselines.
In: International Conference on Medical Image Computing and Computer-Assisted Intervention
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
HU-based Foreground Masking for 3D Medical Masked Image Modeling
HU-based foreground masking, which trains masked image models only on subvolumes with mean intensity above 0.1, improves downstream 3D medical segmentation Dice by 0.3 to 2.9 points over random masking baselines.