Predictive radiomics based ensemble machine learning approach in CT lung nodule diagnosis

Area of Science:

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

Background:

• Computed tomography (CT) is vital for lung cancer diagnosis, a leading cause of cancer mortality.
• Radiomics and machine learning (ML) show promise for identifying lung nodules but require feature selection.
• Accurate classification of pulmonary nodules is essential for timely diagnosis and treatment.

Purpose of the Study:

• To investigate the efficacy of radiomics features for classifying CT-based pulmonary nodules.
• To apply ML techniques for distinguishing benign from malignant lung nodules.
• To identify optimal radiomics features for improved diagnostic accuracy.

Main Methods:

• Utilized the Lung Image Data Consortium (LIDC) database with 1018 CT cancer cases.
• Extracted radiomics features using Wavelet Packet Transform, geometrical features, GLRLM, GLCM, and GLDM.
• Employed boosted and bagged ensemble classification trees for feature selection (BACET, BOCET).
• Classified nodules using ML models including Support Vector Machines, Decision Trees, and various ensemble methods.

Main Results:

• Ensemble Subspace KNN with BACET feature selection achieved the highest AUROC (93.4%), accuracy (88.3%), and F1-score (85.2%).
• FGSVM demonstrated the best sensitivity (97.1%).
• RUSBOCET achieved the best precision (93.4%) and specificity (83.1%).

Conclusions:

• Radiomics analysis combined with ML classifiers effectively classifies pulmonary nodules on CT scans.
• The proposed methodology aids clinicians in decision support for early lung cancer detection.
• Ensemble Subspace KNN and other ML models show significant potential for quantitative CT image analysis in oncology.