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

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Ground-glass opacities (GGOs) on CT scans present a diagnostic challenge in differentiating malignant from benign lung lesions.
• Accurate prediction of GGO malignancy is crucial for timely and appropriate patient management.

Purpose of the Study:

• To compare the diagnostic performance of a linear multivariate statistical regression model against an AI deep learning method for predicting GGO malignancy.
• To evaluate the ability of both methods to distinguish between benign and malignant GGO lesions based on CT scan pixel features.

Main Methods:

• Retrospective analysis of lung CT scans from 47 patients with pure or part-solid GGO lesions.
• Manual segmentation of GGOs and extraction of six image texture features using MaZda software.
• Application of a linear multivariate statistical regression model and a custom-developed AI deep learning model to the extracted features and CT images.

Main Results:

• The multivariate regression model identified two key variables (S(4,4)AngScMom and WavEnLH_s-2) with significant differences between benign and malignant GGOs, achieving 91% sensitivity and 67% specificity (AUC: 0.8).
• The AI deep learning model demonstrated superior performance with 100% sensitivity and 80% specificity (AUC: 0.96).
• Pathology confirmed 32 malignant and 15 benign lesions among the 47 patients.

Conclusions:

• The AI deep learning model significantly outperformed the multivariate statistical regression in predicting GGO malignancy, particularly in sensitivity and specificity.
• These findings suggest the potential of AI in improving the accuracy of GGO lesion characterization.
• Further validation with larger patient cohorts is recommended due to the study's proof-of-concept nature and small sample size.