Pixel level deep reinforcement learning for accurate and robust medical image segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Deep learning methods dominate medical image segmentation but create unsustainable path dependencies with large models.
• Current deep reinforcement learning (DRL) methods for segmentation face challenges like high training costs and uncertain outputs.

Purpose of the Study:

• To introduce a novel Pixel-level Deep Reinforcement Learning model with pixel-by-pixel Mask Generation (PixelDRL-MG) to enhance medical image segmentation accuracy and robustness.
• To break the path dependency of current deep learning models by reducing model parameters and deployment costs.

Main Methods:

• Proposed PixelDRL-MG model utilizing a dynamic iterative update policy for direct segmentation without user interaction or initial masks.
• Introduced a Pixel-level Asynchronous Advantage Actor-Critic (PA3C) strategy, treating each pixel as an agent for iterative state updates.

Main Results:

• PixelDRL-MG demonstrated superior segmentation performance compared to state-of-the-art baselines, particularly at object boundaries.
• The model achieved this with significantly fewer parameters and showed strong performance in low-resource settings (50-shot/100-shot).

Conclusions:

• PixelDRL-MG offers a more accurate, robust, and parameter-efficient solution for medical image segmentation.
• The model's effectiveness in low-resource scenarios makes it suitable for real-world applications.