AI AIDED DISCRIMINATION OF COVID AND PNEUMONIA DATASET WITH DEEP LEARNING BASED 3D SEGMENTATION MODELS

Abstract

The rapid and accurate discrimination between COVID-19 and common pneumonia has become crucial for effective patient management, especially during the global pandemic. This study presents a novel approach using deep learning-based 3D segmentation techniques to differentiate between COVID-19-induced pneumonia and other forms of pneumonia from medical imaging data, specifically computed tomography (CT) scans. The proposed framework aims to assist radiologists and healthcare providers in identifying unique patterns in COVID-19 infections while distinguishing them from common viral or bacterial pneumonia. The core of the method involves the application of a 3D convolutional neural network (CNN) integrated with a V-Net and Res-Net architectures for volumetric segmentation of lung regions affected by infection. By analyzing CT scan volumes, the model can isolate and segment crucial lung abnormalities, such as ground-glass opacities (GGOs), consolidations, and other characteristic features seen in COVID-19 and pneumonia patients. Preprocessing steps, including image normalization, contrast enhancement, and noise reduction, ensure robust input data for model training and testing. The 3D segmentation model is trained on a diverse open publicly dataset comprising 1000 CT scans labeled for COVID-19 and 1000 CT for common pneumonia cases. It utilizes deep learning techniques, including transfer learning, to maximize performance and efficiency, allowing the model to generalize well across varying patient populations. Additionally, the system employs a hybrid classification model that further distinguishes between COVID-19 and non-COVID pneumonia based on the segmented lung regions, using features such as the distribution, volume, and texture of infected areas. Performance evaluation demonstrates that the proposed deep learning model achieves high accuracy, sensitivity, and specificity in distinguishing COVID-19 from common pneumonia. The model outperforms conventional 2D segmentation techniques by leveraging the richer spatial context provided by 3D imaging. In extensive testing, the system attained an accuracy exceeding 96% for ResNet, %86 for SegNet and %74 for UNet models, with notable improvements in reducing both false positives and false negatives. Furthermore, the model's segmentation quality is validated against radiologist annotations, confirming its clinical relevance and potential for real-world applications.