AutoPath: Image-Specific Inference for 3D Segmentation.

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer vision

Background:

• Deep convolutional neural networks (CNNs) have advanced medical image segmentation.
• Residual structures are crucial for CNN-based segmentation.
• 3D residual networks present significant computational challenges for clinical inference.

Purpose of the Study:

• To propose AutoPath, an image-specific inference method for efficient 3D medical image segmentation.
• To reduce the computational burden of 3D segmentation without sacrificing performance.

Main Methods:

• AutoPath dynamically selects residual blocks based on individual input images during inference.
• A policy network is trained using reinforcement learning to optimize block selection.
• The training rewards minimal block usage while maintaining segmentation accuracy.

Main Results:

• Experimental results on a liver CT dataset demonstrate AutoPath's efficiency.
• The approach effectively reduces computational load during inference.
• Satisfactory segmentation performance was achieved.

Conclusions:

• AutoPath offers an efficient solution for 3D medical image segmentation.
• The method addresses the computational limitations of existing 3D residual networks.
• AutoPath shows promise for real-world clinical applications.