Contour Field-Based Elliptical Shape Prior for the Segment Anything Model.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning models like the Segment Anything Model (SAM) are powerful for image segmentation but struggle with efficiently producing elliptical shapes.
• Prior information about elliptical shapes is crucial for improving segmentation accuracy in specific medical and natural image applications.

Purpose of the Study:

• To develop a novel approach for integrating elliptical shape priors into the Segment Anything Model (SAM) for enhanced image segmentation.
• To improve the efficiency and accuracy of SAM in segmenting elliptical regions by incorporating variational methods.

Main Methods:

• Proposed a parameterized elliptical contour field to constrain segmentation results to predefined elliptical contours.
• Utilized a dual algorithm to integrate image features with elliptical and spatial regularization priors.
• Decomposed the SAM into four mathematical subproblems to integrate the variational ellipse prior into a new network structure.

Main Results:

• The proposed method successfully constrains SAM's output to elliptical regions.
• Experimental results show improved segmentation accuracy compared to the original SAM on specific image datasets.
• The integration of elliptical priors enhances the model's ability to segment elliptical objects.

Conclusions:

• The integration of elliptical shape priors via variational methods offers a significant improvement for SAM in segmenting elliptical structures.
• The developed SAM network structure effectively leverages elliptical priors for more accurate and efficient image segmentation.
• This approach holds promise for applications requiring precise segmentation of elliptical objects in medical and natural imaging.