Patch size curriculum accelerates 3D medical segmentation training and boosts Dice scores across 15 tasks

Progressive growing of patch size (PGPS) in 3D medical image segmentation training improves mean Dice by 1.28% across 15 tasks while cutting training time to 89% of standard; in resource-efficient mode, time drops to 44% with matched performance.

• The performance mode improved Dice score across all 15 tasks, with particular benefits for lesion segmentation where foreground-background imbalance is severe.
• The approach is compatible with multiple backbone architectures (UNet, UNETR, SwinUNETR) and reduces performance variance, making model comparisons more reliable.
• The resource-efficient mode maintained Dice scores while cutting training time to 44% of the constant-patch-size baseline.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.