Reinforcement learning-based anatomical maps for pancreas subregion and duct segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Automated pancreas segmentation from medical images is challenging due to anatomical variations.
• Accurate segmentation of pancreatic subregions and the pancreatic duct is crucial for diagnosis and treatment planning.

Purpose of the Study:

• To present a novel framework for segmenting individual pancreatic subregions and the pancreatic duct from 3D CT images.
• To leverage multiagent reinforcement learning (RL) for improved anatomical landmark detection and segmentation accuracy.

Main Methods:

• A multiagent RL network was employed to detect key landmarks of the pancreas and pancreatic duct.
• Nonrigid registration of a pancreatic atlas to detected landmarks generated anatomical probability maps.
• Multilabel 3D U-Net architectures augmented these maps for final segmentation.

Main Results:

• The RL-based framework achieved a mean Dice Similarity Coefficient (DSC) of 0.51 for the pancreatic head, 0.47 for the neck, 0.49 for the body, and 0.49 for the tail, outperforming standard U-Net architectures.
• For the pancreatic duct, the RL-based framework achieved a mean DSC of 0.70, significantly outperforming existing methods.
• Performance was evaluated on 82 CT images for subregions and 37 for ducts.

Conclusions:

• The proposed RL-based segmentation framework demonstrates superior accuracy compared to standard U-Net architectures for pancreas and pancreatic duct segmentation.
• This advancement offers a more robust tool for analyzing pancreatic anatomy in 3D CT imaging.