RLSegNet: An Medical Image Segmentation Network Based on Reinforcement Learning.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Artificial Intelligence

Background:

• Spatial information is crucial for enhancing medical image segmentation.
• 3D convolution effectively utilizes spatial information, but optimizing 2D convolution for spatial information remains a challenge.

Purpose of the Study:

• To propose a novel image segmentation network, RLSegNet, that leverages reinforcement learning (RL) to improve spatial information utilization in 2D convolutions.
• To translate the medical image segmentation process into a sequential decision-making problem.

Main Methods:

• Developed RLSegNet, a U-shaped network comprising a feature extraction network, a Mask Prediction Network (MPNet), and an up-sampling network with a cascade attention module.
• Redesigned state, action, and reward mechanisms within the RL framework to optimize the segmentation decision-making process.
• Applied the RLSegNet for brain tumor segmentation on the BRATS 2015 dataset.

Main Results:

• RLSegNet demonstrated superior performance in brain tumor segmentation compared to existing state-of-the-art methods.
• The cascade attention module effectively generated weighted feature masks, guiding the up-sampling network to focus on relevant regions.
• The RL-based approach successfully translated segmentation into an effective decision-making process.

Conclusions:

• The proposed RLSegNet effectively utilizes spatial information within 2D convolutions through a reinforcement learning framework.
• RLSegNet achieves improved segmentation performance, particularly for brain tumor segmentation tasks.
• This work presents a promising new direction for medical image segmentation by integrating reinforcement learning.