WCLC 2025 - Posters & ePosters-P1.04.46 Predicting Malignancy in Ground-Glass Opacity Using Multivariate Regression and Deep Learning Models: A Proof-Of-Concept

P1.04.46 Predicting Malignancy in Ground-Glass Opacity Using Multivariate Regression and Deep Learning Models: A Proof-Of-Concept

Back to course

Pdf Summary

This proof-of-concept study aims to improve the non-invasive differentiation between benign and malignant ground-glass opacities (GGOs) detected on lung CT scans—a diagnostic challenge in lung cancer screening. The retrospective analysis involved 47 patients with confirmed pure or part-solid GGOs from an initial cohort of 72. GGOs were manually segmented, and texture features were extracted from CT images using MaZda software, which evaluates six image texture feature groups. These features were analyzed via multivariate statistical regression to discriminate malignancy, while a parallel cloud-based deep learning model was applied to the same images for comparison.<br /><br />Univariate analysis identified 26 texture variables showing significant or near-significant differences between benign and malignant lesions; multivariate regression narrowed this to two independent variables (S(4,4)AngScMom and WavEnLH_s-2). Using these variables, the regression model achieved a sensitivity of 91% and specificity of 67%. By contrast, the deep learning model demonstrated superior performance with 100% sensitivity and 80% specificity, indicating its enhanced capability to detect malignant GGOs accurately.<br /><br />These results highlight the potential of deep learning models over traditional multivariate regression in characterizing GGOs, which could lead to improved early lung cancer diagnosis. However, the study recognizes its limitations due to the modest sample size, suggesting that validation in larger patient cohorts is necessary. The research builds on prior radiomic and AI approaches for GGO evaluation and contributes new evidence supporting the integration of advanced computational tools into thoracic imaging workflow.<br /><br />In conclusion, this study demonstrates that AI-driven deep learning models can outperform conventional statistical methods in predicting malignancy in lung GGOs, offering promising avenues for more precise, non-invasive lung cancer diagnostics. The findings encourage further exploration and refinement with expanded datasets.

Asset Subtitle

Abed Agbarya

Meta Tag

Speaker Abed Agbarya

Topic Screening and Early Detection

Keywords

ground-glass opacities

lung cancer screening

CT imaging

texture analysis

MaZda software

multivariate regression

deep learning model

sensitivity and specificity

non-invasive diagnosis

radiomics