Improving malignancy prediction through feature selection informed by nodule size ranges in NLST

Area of Science:

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

Background:

• Computed tomography (CT) is crucial for Non-Small Cell Lung Cancer (NSCLC) diagnosis and treatment.
• Current computer-aided diagnosis (CAD) models for lung nodule classification rely on selected image features.
• Nodule size variation suggests potential differences in radiomic descriptors between benign and malignant cases.

Purpose of the Study:

• To investigate automated selection of image features tailored to specific nodule size ranges.
• To enhance the accuracy of classifying malignant and benign lung nodules in CT scans.
• To evaluate the impact of nodule size-informed feature selection on diagnostic performance.

Main Methods:

• Utilized the National Lung Screening Trial (NLST) dataset, comprising 261 training and 237 testing cases.
• Split datasets into three subsets based on the longest nodule diameter (LD) for size-specific analysis.
• Implemented automated feature selection and evaluated classification performance, including the effect of oversampling the minority cancer class.

Main Results:

• Accuracy improved from 74.68% to 81.01%, and AUC improved from 0.69 to 0.79, by splitting datasets based on nodule size.
• When AUC was prioritized, accuracy increased from 72.57% to 77.5%, and AUC rose from 0.78 to 0.82.
• Oversampling the minority cancer class further boosted performance in specific cases, from 0.82 to 0.87.

Conclusions:

• Automated feature selection informed by nodule size significantly improves the accuracy and AUC of lung nodule classification in CT scans.
• This size-stratified approach offers a more refined method for CAD systems in NSCLC diagnosis.
• The findings highlight the importance of considering nodule size variability for enhanced diagnostic accuracy.