Reproduce “Primitive Geometry Segment Pre-training for 3D Medical Image Segmentation” on OpenResearch

About this paper

Ryu Tadokoro, Ryosuke Yamada, Kodai Nakashima, Ryo Nakamura, Hirokatsu Kataoka

A new pre-training method, PrimGeoSeg, employs automatically generated synthetic 3D geometric objects to improve medical image segmentation models. The method achieves an average Dice score of 82.0% on the BTCV dataset, outperforming learning from scratch by 3.7%, and demonstrates high data efficiency by matching full-data scratch performance with only 30% of real training data.

Key insights

• Pre-training with volumetric shapes is essential, yielding a 7.78 Dice point improvement over planar shapes, highlighting the critical role of true 3D structural information.
• Both inter-class diversity (e.g., 32 distinct shape classes) and intra-class variability (Instance Augmentation) in synthetic data are crucial, leading to 2.4 and 3.31 Dice point improvements respectively.
• Incorporating object overlap during synthetic data generation improves segmentation accuracy by 0.62 Dice points, suggesting that mimicking complex, ambiguous anatomical boundaries enhances learning.

Summary by alphaXiv.