Predicting Lung Adenocarcinoma Invasiveness via Machine Learning on CT GGNs

Overview

This study developed a machine learning framework integrating CT-based radiomics and clinical data to non-invasively predict the invasiveness of lung adenocarcinoma presenting as ground-glass nodules (GGNs). Using a multicenter retrospective cohort, the model aims to assist clinicians in preoperative decision-making by improving the accuracy of malignancy assessment.

Background

Lung adenocarcinoma is the most common subtype of lung cancer and often presents as ground-glass nodules on CT scans. Early detection is critical due to the poor prognosis of advanced lung cancer and the limitations of current screening methods, including subjective radiological assessment and insufficient biomarker sensitivity. Radiomics offers quantitative imaging features that may reveal tumor biology beyond traditional imaging, potentially improving the evaluation of nodule invasiveness. This study addresses the challenge of distinguishing invasive adenocarcinoma from less aggressive lesions preoperatively by combining radiomics with deep learning and clinical parameters.

Data Highlights

The study included 312 patients from the Second Hospital of Lanzhou University as the primary cohort and 45 patients from Yantai Yuhuangding Hospital as the validation cohort. Inclusion criteria mandated CT imaging within one week before surgery with slice thickness ≤1.25 mm and nodules less than 3 cm. CT scans were performed using multiple systems with standardized parameters (120 kVp, 150-200 mA, 5 mm axial thickness, 1.25 mm reconstruction thickness). Imaging features such as nodule size, solid component diameter, and morphological characteristics were assessed by experienced radiologists blinded to pathology.

Key Findings

• A machine learning model combining CT radiomics and clinical features can predict lung adenocarcinoma invasiveness in GGNs preoperatively.
• Radiomics captures quantitative features like shape, texture, and density that are not discernible by traditional imaging assessment.
• Subjective radiological evaluation alone is limited by overlapping features between minimally invasive and invasive adenocarcinomas, especially in pure ground-glass nodules.
• Integration of deep learning-derived features enhances the predictive accuracy beyond conventional radiomics.
• Multicenter data and standardized CT protocols support the generalizability of the model.

Clinical Implications

This approach provides a non-invasive, objective tool to assist clinicians in distinguishing invasive from non-invasive lung adenocarcinoma in patients with GGNs, potentially guiding personalized surgical planning and treatment strategies. Early and accurate prediction of invasiveness may reduce overtreatment and improve patient outcomes by tailoring interventions to tumor biology.

Conclusion

Combining radiomics and machine learning with clinical CT data offers a promising method for preoperative prediction of lung adenocarcinoma invasiveness in ground-glass nodules, facilitating more precise and individualized clinical decision-making.

Related Resources & Content

1. Global Cancer Statistics and Lung Cancer Mortality
2. Radiomics and Lung Cancer Screening Advances
3. Clinical CT Imaging Protocols and Feature Analysis