Edge-Sensitive Left Ventricle Segmentation Using Deep Reinforcement Learning.

Area of Science:

• Medical image analysis
• Computer vision
• Artificial intelligence

Background:

• Deep reinforcement learning (DRL) is widely used in computer vision but less explored for image segmentation, particularly left ventricle (LV) segmentation.
• Existing reinforcement learning methods for segmentation often use thresholding, leading to inaccuracies due to sensitivity.

Purpose of the Study:

• To develop a novel DRL agent for accurate LV segmentation by imitating the human segmentation process.
• To address the limitations of threshold-based reinforcement learning methods in LV segmentation.

Main Methods:

• Formulated LV segmentation as a Markov decision process optimized via DRL.
• Designed a DRL agent with two neural networks: First-P-Net for initial edge point detection and Next-P-Net for successive edge point localization.
• Achieved closed segmentation results by iteratively identifying edge points.

Main Results:

• The proposed DRL model outperformed previous reinforcement learning methods for LV segmentation.
• Achieved comparable performance to deep learning baselines on the ACDC 2017 and Sunnybrook 2009 datasets.
• Demonstrated higher F-measure accuracy than deep learning methods when trained with limited samples.

Conclusions:

• The novel DRL agent effectively performs LV segmentation, offering an improvement over existing reinforcement learning techniques.
• The model shows promise for accurate cardiac image segmentation, particularly in data-scarce scenarios.