Mesh Regression Based Shape Enhancement Operator Designed for Organ Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Computational anatomy

Background:

• Organ delineation is crucial for medical diagnosis and treatment planning.
• Current neural network methods achieve high segmentation metrics but struggle with indistinct boundaries due to ignoring anatomical priors.
• Anatomical information is vital for accurate organ segmentation, mimicking expert radiologist approaches.

Purpose of the Study:

• To propose a mesh regression-based shape enhancement operator to refine organ segmentation masks.
• To address the limitation of indistinct boundaries in current pixel-based segmentation methods.
• To integrate anatomical priors into the segmentation refinement process.

Main Methods:

• A mesh regression operator is proposed, modeling contour refinement as a mesh vertex regression task.
• The operator refines coarse segmentation masks from any existing model.
• Graph convolutional neural networks predict vertex displacements using fast point feature histograms, evolving the mesh for improved segmentation contours.

Main Results:

• The proposed operator significantly improves organ segmentation performance on public datasets for liver and pancreas delineation.
• Validation demonstrated the effectiveness of the two-stage segmentation pipeline in refining results based on geometrical characteristics.
• The shape enhancement operator proved to be plug-and-play, compatible with various backbone segmentation models.

Conclusions:

• The mesh regression-based shape enhancement operator effectively refines organ segmentation by incorporating anatomical priors.
• This approach overcomes the limitations of pixel-based classification, yielding more accurate segmentation contours.
• The method shows significant promise and application prospects in medical image analysis and segmentation.