Deep Reinforcement Learning for Weakly-Supervised Lymph Node Segmentation in CT Images.

Area of Science:

• Medical Imaging Analysis
• Computational Pathology
• Artificial Intelligence in Oncology

Background:

• Accurate lymph node segmentation is crucial for disease progression assessment and therapeutic evaluation.
• Challenges in lymph node segmentation include morphological variations and the difficulty of obtaining manual annotations.
• Response Evaluation Criteria in Solid Tumors (RECIST) annotations are often available and can serve as a basis for weakly-supervised learning.

Purpose of the Study:

• To develop a weakly-supervised deep reinforcement learning model for automated lymph node segmentation.
• To leverage readily available RECIST annotations for training segmentation models.
• To simultaneously optimize lymph node bounding boxes and segmentation results.

Main Methods:

• Proposed a deep reinforcement learning-based lymph node segmentation (DRL-LNS) model.
• Utilized RECIST annotations to generate pseudo ground truths for training a U-Net segmentation network.
• Implemented a deep reinforcement learning model where the segmentation and policy networks interact to refine bounding boxes and segmentation.

Main Results:

• The DRL-LNS model achieved a mean Dice Similarity Coefficient (DSC) of 77.17% and a mean Intersection over Union (IoU) of 64.78% on a public thoracoabdominal CT dataset.
• The model was evaluated against three established image segmentation networks.
• The DRL-based bounding box prediction strategy demonstrated superior performance compared to label propagation.

Conclusions:

• The proposed DRL-LNS model achieves state-of-the-art performance for weakly-supervised lymph node segmentation.
• Leveraging RECIST annotations offers a viable approach for training segmentation models when detailed annotations are scarce.
• The DRL-LNS model shows significant potential for improving quantitative disease assessment in clinical practice.