A Reinforcement Learning Framework for Lung Segmentation of COVID-19 and Pneumonia Affected Chest X-Ray Image

Deep CNN models are generally designed based on the trial-and-error method. This manner needs a significant amount of time and memories. To tackle this problem, we proposed a reinforcement learning method for biomedical image segmentation where all the hyperparameters of the model will be selected automatically. In this paper, we segmented lungs from the Pneumonia, COVID-19, and Healthy X-ray images by merging Reinforcement Learning and Deep CNN. At first, use different image processing techniques on input images. Later, use pre-trained VGG-19 as a feature extractor. In this approach, the agent is trained based on the maximum reward function and creates the U-Net architecture using the policy gradient RL technique. This research is conducted on a dataset of 3,877 images. This approach achieved an accuracy of 98.03% and a Dice index of 0.9543 on the test dataset.