A Fully Automatic Approach for COVID-19 Diagnosis in CT Imaging: Integrating Lung Segmentation, Fine-Tuning and Grad-CAM Visualization

• Matheus A. dos Santos UFC
• Iágson Carlos L. Silva UFC
• Lucas de O. Santos UFC
• Elizângela de S. Rebouças IFCE / UMBC
• Pedro Pedrosa Rebouças Filho IFCE / UMBC
• Houbing H. Song UMBC

Resumo

The pandemic caused by the COVID-19 virus has highlighted the need for efficient and automated medical diagnostic tools to assist healthcare professionals in the fast and accurate identification of the disease. This study introduces a fully automatic approach for classifying COVID-19 in thoracic computed tomography (CT) images from the SARS-CoV-2 CT-scan dataset, employing deep learning, automatic lung segmentation, and fine-tuning models to the specific problem. Previous transfer learning experiments revealed that MobileNet emerged as the most effective architecture for feature extraction. Integrating Detectron2 for lung segmentation and subsequent fine-tuning of the selected MobileNet model significantly improved classification performance. Visual analysis of model interpretability using Grad-CAM demonstrated that with segmented images and the refined model, the network focused on lung regions, enhancing understanding of the model diagnosis. As a result, the proposed fully automatic method achieved superior evaluation metrics compared to other methods in the literature, reaching values of 99.60% for accuracy, precision, and F1-Score, along with 99.59% for recall and 99.20% for the Matthews Correlation Coefficient (MCC).