TeTrIS: Template Transformer Networks for Image Segmentation With Shape Priors.

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Neural networks are powerful tools for image segmentation.
• Incorporating prior knowledge, such as shape, can enhance segmentation performance.
• Existing methods may struggle with complex shapes or limited data.

Purpose of the Study:

• To introduce and compare novel methods for incorporating shape prior information into neural network-based image segmentation.
• To demonstrate the benefits of shape priors in improving segmentation accuracy and reducing false positives.
• To apply these methods to synthetic data and real-world medical imaging tasks.

Main Methods:

• Developed template transformer networks using spatial transformer networks for end-to-end shape deformation.
• Integrated shape priors as an additional input channel for state-of-the-art methods like Fully Convolutional Networks and U-Net.
• Utilized soft partial volume segmentation to avoid discretization artifacts.

Main Results:

• Template transformer networks explicitly enforce shape priors and provide soft segmentations.
• Adding shape priors as an input channel significantly reduced false positives in pixel-wise classification.
• Demonstrated improved segmentation of coronary lumen structures in cardiac CT data.

Conclusions:

• Incorporating shape prior information is beneficial for neural network-based image segmentation.
• The proposed methods offer effective ways to integrate shape priors, leading to more accurate and robust segmentations.
• These techniques show promise for various medical imaging applications requiring precise structural segmentation.