Abstract
Background: The widespread use of high-resolution computed tomography (HRCT) in lung cancer screening has allowed for an increased detection rate of ground-glass nodules (GGNs) in the lung. Hence, obtaining the correct clinical diagnosis of benign and malignant GGNs has become crucial. Objectives: Most artificial intelligence and computer-aided diagnosis (AI-CAD) systems for the classification of pulmonary GGNs fail to extract CT features. This study used HRCT and AI to analyze the CT features of GGNs to improve the prediction of benign and malignant pulmonary GGNs. Patients and Methods: This case-control study was performed on a malignant group consisting of patients and a benign group consisting of controls. A total of 204 patients with GGNs were recruited and divided into 2 groups according to their pathological results. Group A consisted of 69 cases with precursor glandular lesions (atypical adenomatous hyperplasia [AAH] and adenocarcinoma in situ [AIS]), inflammatory nodules, and benign nodules. Group B consisted of 135 cases with invasive lesions (minimally invasive adenocarcinoma [MIA], invasive adenocarcinoma [IAC], and other malignant lesions). Various CT features were compared between the 2 groups. The diagnostic efficacy of an AI-CAD system and radiologists’ reports for benign and malignant nodules were analyzed. A multivariate logistic regression analysis was performed to determine independent predictors of malignant GGN. A model that combined the AI system and manual extraction of radiological features was constructed. The receiver operating characteristic (ROC) curve was used to evaluate the diagnostic efficiency of the model. Results: Significant differences were found between malignant and benign groups according to the following 7 CT features: The GGN size (long and short diameters), vacuole sign, air bronchogram sign, vascular convergence sign, vascular perforator sign, interlobular septal obstruction sign, and spiculation (P < 0.05). The volume and mean CT values of precursor glandular lesions of the lungs were significantly different from those of invasive lesions (P < 0.05). The logistic regression model showed that the sensitivity and specificity of the AI system in diagnosing malignant groups were 0.756 and 0.696, respectively. The sensitivity and specificity of radiologists’ reports in diagnosing the malignant groups were 0.726 and 0.783, respectively. The combination of the 2 had a sensitivity of 0.768 and a specificity of 0.793. Conclusion: Prediction of the nature of GGNs based on CT features, including the vacuole sign, vascular perforator sign, and interlobular septal obstruction sign, were relatively accurate for a preliminary diagnosis. The AI system had a poorer diagnostic accuracy for GGNs than radiologists’ reports of CT images. The combination of AI and radiologists’ reports showed the highest diagnostic efficacy.