Efficient pulmonary nodules classification using radiomics and different artificial intelligence strategies.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Machine Learning

Background:

• Pulmonary nodule classification is crucial for early lung cancer detection.
• Accurate classification aids in treatment planning and patient outcomes.
• Developing efficient AI models for nodule analysis is an ongoing research area.

Purpose of the Study:

• To explore and develop artificial intelligence (AI) approaches for efficient pulmonary nodule classification using CT scans.
• To compare the performance of traditional machine learning (ML) methods with deep learning (DL) techniques.
• To identify the most effective AI models for pulmonary nodule classification.

Main Methods:

• Utilized the LIDC-IDRI dataset comprising 1007 nodules from 551 patients.
• Preprocessed nodule images (64x64 PNG) to remove non-nodular structures.
• Extracted texture features (Haralick, LBP) for ML, followed by PCA for feature selection.
• Developed a simple Convolutional Neural Network (CNN) and applied transfer learning with pre-trained models (VGG-16/19, DenseNet-121/169, ResNet) for DL.

Main Results:

• Statistical ML achieved optimal AUROC of 0.885 (Random Forest) and accuracy of 0.819 (SVM).
• Deep learning models demonstrated superior performance: DenseNet-121 reached 90.39% accuracy.
• Highest AUROC (96.0%) achieved with simple CNN; best sensitivity (90.32%) with DenseNet-169; best specificity (93.65%) with DenseNet-121 and ResNet-152V2.

Conclusions:

• Deep learning with transfer learning offers significant advantages over statistical ML for pulmonary nodule classification.
• DenseNet-121 and SVM models exhibited the best performance among the evaluated methods.
• Future improvements are possible with larger datasets and 3D volumetric analysis for enhanced lung cancer diagnosis.