Improved U-Net Model Based on Deep Q-Network Optimisation for COVID-19 Lung CT Image Segmentation

The emergence of the novel coronavirus (2019-nCoV) presented a significant challenge to healthcare systems worldwide. Computed tomography (CT) scans were instrumental in delineating infected areas, facilitating the rapid identification and isolation of individuals with the virus. However, the accuracy of image segmentation is constrained by the complexity of CT images of the Coronavirus disease 2019 (Covid-19) and the paucity of labelled data. In light of the aforementioned limitations, this paper proposes the DQDPU-Net model as a means of enhancing the efficacy of lung lesion segmentation from CT images of the lungs infected with the SARS-CoV-2 virus. This approach entails the integration of deep learning and reinforcement learning techniques. Firstly, the U-Net structure is enhanced through the introduction of a dilated convolution block (DC) with residuals, which expands the sensory field to obtain more contextual information. Secondly, the PPM and MSA modules are integrated to improve the extraction of detail information from the feature maps. Lastly, the EDPath block is developed to address the issue of high-resolution information loss. This paper further introduces the DQN algorithm, which takes the segmentation network as the environment. The intelligent body improves the segmentation accuracy and stability of the model by optimising the action strategy. The experimental results demonstrate that the DQUNet method exhibits superior performance compared to existing techniques in terms of DICE, IOU, and SEN metrics.