Machine Learning and Feature Selection Methods for Disease Classification With Application to Lung Cancer Screening...

Area of Science:

• Radiology
• Machine Learning
• Oncology

Background:

• Lung cancer screening has increased awareness of risks and benefits.
• The National Lung Screening Trial (NLST) demonstrated reduced mortality but had a high false positive rate (>94%).
• Improving the accuracy and reducing false positives in lung nodule classification is crucial.

Purpose of the Study:

• To evaluate machine learning classifiers for predicting lung nodule malignancy.
• To assess the ability of these models to reduce the false positive rate.
• To identify optimal classifiers and feature selection methods for lung nodule classification.

Main Methods:

• Utilized 416 quantitative imaging biomarkers from CT scans of 200 patients with verified cancerous or benign lung nodules.
• Employed various linear, nonlinear, and ensemble machine learning models.
• Applied feature selection methods, including linear combination and correlation, to classify nodule status.

Main Results:

• Elastic net and support vector machine models, with specific feature selection, achieved the highest performance (average cross-validation AUC near 0.72).
• Random forest and bagged trees performed less effectively (AUC near 0.60).
• The best-performing models achieved a false positive rate of approximately 30%, a significant reduction compared to the NLST.

Conclusions:

• Radiomic biomarkers combined with machine learning offer a promising approach for lung nodule classification.
• These methods demonstrate potential for accurate tumor classification while substantially lowering false positive rates.
• This technology could enhance diagnostic accuracy and improve patient outcomes in lung cancer screening.