Orthogonal Seeding Boosts 3D Medical Image Segmentation by 22%
Orthogonal seeding cuts organ segmentation errors by 53% vs. single-axis methods...
Deep Dive
A new inference-time strategy for 3D organ segmentation in CT/MRI uses three orthogonal seed slices—axial, coronal, and sagittal—instead of a single axial seed, then fuses their propagated labels with a simple label-free rule. Tested on a multi-organ CT cohort with the Sli2Vol slice-propagation model, this approach improves Dice by 21.9%, Normalized Surface Dice by 25.5%, and reduces Average Hausdorff Distance by 53.5% over the
Key Points
- Orthogonal Seeding uses three axial/coronal/sagittal seeds to align 3D organ segmentation, improving Dice score by 21.9%
- Works with existing Sli2Vol models—no retraining needed; gains come from inference-time seed geometry
- Reduces surface-distance errors (e.g., Hausdorff Distance) by 53.5% vs. single-axis methods in CT/MRI cohorts
Why It Matters
Could dramatically reduce manual annotation costs for 3D medical imaging while improving AI-assisted diagnostics.