Detection and measurement of multilevel COVID-19 infection using gamma correction and features extracted by CNN enhanced with xgboost from CT scan images

The field of radiation diagnosis is undergoing rapid advancements, propelled by innovative technologies such as artificial intelligence (AI), computer vision, and sophisticated medical imaging techniques. In this dynamic landscape, computed tomography (CT) imaging has emerged as a prominent method, especially in the identification of clinical changes associated with both COVID-19 and lung tumors. The utilization of chest X-rays, various medical imaging techniques, and CT scans has been proposed as a practical and efficient means for the rapid and accurate diagnosis of COVID-19. Notably, CT scans have demonstrated high sensitivity in detecting the presence of COVID-19, even in cases where false negatives may occur under extreme circumstances. However, the analysis of X-rays and CT images by specialists to determine COVID-19 positivity is a time-consuming and challenging process. To address these challenges and improve the diagnostic process, the study addresses the application of convolutional neural networks (CNNs). These CNNs play a crucial role in accurately categorizing images while filtering out irrelevant elements, a feat achieved through transfer learning. In transfer learning, pre-existing methods gather essential characteristics from extensive datasets like ImageNet, effectively applying them to new tasks. Despite the advancements in machine learning techniques, the efficacy of CNNs still hinges on the extraction of relevant attributes from CT images. Therefore, the study suggests that the use of chest CT scans and X-ray images holds significant potential for detecting COVID-19, and this potential can be further enhanced by employing pre-trained CNNs such as DenseNet, ResNet, and VGG-16. The results from two distinct datasets, namely Covid Data 1 and Covid Data 2, along with various models including CNN net and XGBoost classifier, have been compared and analyzed. Across both datasets, XGBoost consistently outperforms the CNN net in terms of accuracy, regardless of the image size and training settings. Covid Data 1 generally exhibits higher accuracies compared to Covid Data 2 across all models and settings tested. Specifically, XGBoost achieves the highest accuracy, reaching up to 97.94% for Covid Data 1 and 97.76% for Covid Data 2, when trained with GC (ϒ=1.5) and using image sizes of 128x128 and 256x256 pixels. On the other hand, the CNN net achieves lower accuracies, with the highest being 79.54% for Covid Data 1 and 79.74% for Covid Data 2, both without GC and using image sizes of 256x256 pixels. Therefore, based on the provided results, the XGBoost classifier appears to be the superior choice for classification tasks, with Covid Data 1 being the preferred dataset due to its consistently higher accuracies across models and settings. While the study makes significant strides toward the therapeutic application of this model for diagnostic purposes, it acknowledges the need for further research to fully realize its potential. Future investigations are expected to explore additional designs utilizing gamma correction and pre-processing tools. To train more complex platforms, considerations for model training options that allow for a larger GPU memory limit are suggested. This could involve incorporating inception modules and multiple dense layers for enhanced capabilities. The next area of research outlined in the study focuses on the classification of various ailments, aiming to determine whether specific architectural styles are more effective in treating certain conditions. Exploring performance differences related to different illnesses could provide insights into the underlying causes. Additionally, alternative data splitting methods, such as k-fold data splitting, are suggested to enhance the analysis process, providing a more comprehensive and robust evaluation of the model's performance.