A Framework for Parallel Segmentation of Lung Nodule Images Based on Reinforcement Learning Enhancement With Multiple Agents

Abstract

Aiming at the problem of lung nodule image segmentation accuracy, a segmentation framework combining TransU-Net and multiagent reinforcement learning was proposed. The powerful global modeling ability of the transformer is adopted to obtain long-distance dependent features, and then combined with the encoder–decoding structure of U-Net to achieve high-precision image restoration and detail preservation, thereby generating a segmentation probability graph with more semantic consistency. Then, the probability map is dynamically divided into multiple overlapping regions, which are processed by the agent in parallel. Each agent makes autonomous decisions based on local features and shares boundary information and global rewards through graph neural networks to achieve information collaboration and strategy collaboration among multiple agents. This mechanism supports the agent to continuously perceive the local state in the graph structure, exchange neighborhood features and respond to the overall feedback, and achieve dynamic collaborative optimization in the constantly updated strategy, thereby improving the segmentation accuracy. The self-attention mechanism is introduced to enhance global perception, and a sharing strategy network is designed to optimize the integration of local and global information. The experiments on the LIDC-IDRI and LUNA16 datasets with complex morphological structures and high fuzzy boundaries show that the Dice coefficient of the proposed method reaches 91.03 and the IoU reaches 83.15, which are significantly better than the existing methods and show good generalization ability.