Malignancy in Ground-Glass Opacity Using Multivariate Regression and Deep Learning Models: A Proof-of-Concept Study

Background/Objectives: Ground-glass opacity (GGO) refers to areas of increased lung opacity on computed tomography (CT) scans. Distinguishing malignant from benign lesions using CT scans remains significantly challenging. This study aims to compare the performances of a linear multivariate statistical regression and an AI deep learning method in their abilities to predict GGO malignancy, given a set of pixel features extracted from CT scans. Methods: This retrospective study investigated patients from the Carmel Medical Center with findings of GGO nodules in their lung CT scans. Forty-seven consecutive patients were found to have either pure or part-solid GGO lesions, as defined by two independent radiologists. After manually segmenting the GGOs in the CT scans, pixel features were extracted using the MaZda software package, which analyzes six different image texture features. These textural variables were then compiled as input for the multivariate statistical regression. Additionally, an AI deep learning method, developed by our group and hosted on the cloud, was applied to the CT images containing the GGOs. Results: Among the 47 patients, 32 were diagnosed by pathology with malignant lesions and 15 with benign findings. Using the multivariate statistical regression, we identified 19 variables with statistically significant or near-significant differences through univariate analysis. In subsequent multivariate analyses, two independent variables that could distinguish between benign and malignant GGO lesions were identified: S(4,4)AngScMom (p = 0.012) and WavEnLH_s-2 (p = 0.008). The regression formula based on these two variables yielded a sensitivity of 91% and a specificity of 67% AUC: 0.8 (95% CI: [0.65, 0.94]). The AI deep learning model demonstrated a sensitivity of 100% and a specificity of 80% AUC: 0.96 (95% CI: [0.86, 1.00]). Conclusions: This proof-of-concept study demonstrates the superior performance of the AI deep learning model compared to the multivariate statistical regression, particularly in terms of sensitivity and specificity. However, given the small sample size, these results could potentially change with larger patient cohorts.