Computer-aided lung nodule recognition by SVM classifier based on combination of random undersampling and SMOTE

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Computer-aided detection/diagnosis (CAD) systems for lung cancer rely on accurate classification of regions of interest (ROI) for lung nodule detection.
• Unbalanced datasets significantly hinder the performance of classification models in these systems.

Purpose of the Study:

• To enhance the generalization ability and classification performance of lung nodule recognition systems facing unbalanced datasets.
• To propose a novel Support Vector Machine (SVM) classifier integrated with resampling techniques.

Main Methods:

• Implemented a combined resampling strategy using random undersampling (RU) and SMOTE (Synthetic Minority Over-sampling Technique) to balance training data.
• Extracted eight features, including both 2D and 3D, for training and classification of pulmonary nodules.
• Utilized a Support Vector Machine (SVM) classifier with the balanced dataset.

Main Results:

• The proposed RU-SMOTE-SVM classifier demonstrated superior performance compared to other classifiers across various training dataset sizes.
• Achieved the highest classification accuracy, with an average exceeding 92.94% for pulmonary nodule classification.
• The resampling combination effectively balanced training samples, removed noise and duplicate data, and retained essential information.

Conclusions:

• The RU-SMOTE-SVM approach significantly improves the performance of SVM algorithms for pulmonary nodule classification in unbalanced datasets.
• This method enhances effective data utilization and increases the generalization ability of lung nodule detection systems.
• The findings suggest a promising direction for developing more accurate and robust CAD systems for lung cancer diagnosis.