Optimized Lung Nodule Classification Using CLAHE-Enhanced CT Imaging and Swin Transformer-Based Deep Feature...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Lung cancer is a leading cause of cancer-related mortality worldwide.
• Early detection of lung cancer significantly improves patient survival rates.
• Computed Tomography (CT) scans are crucial for diagnosing lung cancer.

Purpose of the Study:

• To develop and evaluate a hybrid computer-aided diagnosis (CAD) pipeline for accurate lung cancer classification using CT images.
• To investigate the effectiveness of combining various image preprocessing techniques, deep learning models for feature extraction, and classical machine learning classifiers.
• To enhance the interpretability and reduce overfitting in lung cancer classification models.

Main Methods:

• A hybrid CAD pipeline was designed, integrating ten image preprocessing methods and ten pretrained deep learning models (CNNs, Transformers) for feature extraction.
• Four classical machine learning classifiers were employed for the final classification task.
• The pipeline decoupled feature extraction from classification, allowing for 400 distinct model configurations to be evaluated.
• The approach was validated on the Lung Image Database Consortium and Image Database Resource Initiative dataset (1018 CT scans).

Main Results:

• The optimal pipeline combined Contrast Limited Adaptive Histogram Equalization, Swin Transformer feature extraction, and eXtreme Gradient Boosting.
• This best-performing configuration achieved a high accuracy of 95.8% in classifying malignant versus benign lung nodules.
• The dataset comprised 6568 malignant and 4849 benign images from 1018 thoracic CT scans.

Conclusions:

• The proposed hybrid CAD pipeline demonstrates significant potential for accurate and reliable lung cancer classification from CT scans.
• Decoupling feature extraction and classification offers improved interpretability and robustness compared to end-to-end deep learning systems.
• This approach represents a promising advancement in AI-driven tools for early lung cancer detection and diagnosis.