Abdominal organ segmentation via deep diffeomorphic mesh deformations

Area of Science:

• Medical imaging analysis
• Computational anatomy
• Deep learning for medical applications

Background:

• Abdominal organ segmentation in CT/MRI is crucial for surgical planning and navigation.
• High variability in organ shape, size, and position presents segmentation challenges.
• Template-based methods offer direct mesh reconstruction but struggle with generalization.

Purpose of the Study:

• To assess the generalization capabilities of template-based deep learning methods for abdominal organ segmentation.
• To develop an improved method for joint segmentation of liver, kidney, pancreas, and spleen.
• To enhance segmentation accuracy through a novel architecture and post-processing technique.

Main Methods:

• Employed template-based mesh reconstruction for joint segmentation of four abdominal organs.
• Developed a novel deep diffeomorphic mesh-deformation architecture (UNetFlow).
• Implemented an improved training scheme and a registration-based post-processing step.

Main Results:

• Previous methods showed limited generalization to different organ geometries and small datasets.
• UNetFlow demonstrated good generalization across all four organs and adaptability to new data.
• Post-processing improved the accuracy of both voxel and mesh segmentation outputs.

Conclusions:

• UNetFlow significantly improves the generalization of template-based abdominal organ segmentation.
• The method is robust and can be fine-tuned for clinical deployment.
• Combined deep learning and registration offers a powerful approach for medical image segmentation.