Implicit Anatomical Rendering for Medical Image Segmentation with Stochastic Experts.

Area of Science:

• Medical Imaging
• Computer Vision
• Deep Learning

Background:

• Medical image segmentation requires integrating semantic content and anatomical features.
• Deep learning methods show promise but struggle with boundary details due to grid-based convolutions.
• Implicit neural representations offer advantages over discrete grid-based methods for complex signals.

Purpose of the Study:

• To introduce MORSE, a novel implicit neural rendering framework for medical image segmentation.
• To address limitations of grid-based convolutions in capturing high-frequency boundary details.
• To improve the accuracy and robustness of medical image segmentation.

Main Methods:

• Formulating medical image segmentation as an end-to-end rendering problem.
• Utilizing implicit neural representation for continuous coordinate-based feature alignment.
• Employing a Mixture-of-Expert (MoE) approach with stochastic gating for multi-scale feature optimization.
• Refining boundary regions by adaptively aggregating coordinate-based point features.

Main Results:

• MORSE demonstrates consistent performance improvements when integrated with various medical segmentation backbones.
• The framework achieves competitive results in both 2D and 3D supervised medical image segmentation tasks.
• MORSE effectively refines ambiguous boundary regions, leading to enhanced segmentation accuracy.

Conclusions:

• MORSE offers a powerful and generic framework for improving medical image segmentation.
• Implicit neural rendering provides a superior alternative to grid-based methods for capturing intricate details.
• The proposed method enhances existing segmentation backbones, showing broad applicability and effectiveness.