Ensemble learning driven Kolmogorov-Arnold Networks-based Lung Cancer classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early lung cancer (LC) detection is critical for reducing mortality rates.
• Traditional diagnostic methods face challenges in accurately identifying LC from medical imaging.
• Positron Emission Tomography/Computed Tomography (PET/CT) offers valuable metabolic and anatomical data for LC diagnosis.

Purpose of the Study:

• To develop a robust and accurate LC detection model using enhanced deep learning techniques.
• To leverage PET/CT imaging for comprehensive LC feature extraction and classification.
• To improve the interpretability and efficiency of LC diagnosis in clinical practice.

Main Methods:

• Enhanced MobileNet V3 and LeViT models were used for feature extraction from PET/CT images.
• A weighted sum feature fusion technique combined extracted features.
• Kolmogorov-Arnold Networks (KANs) with spline functions (linear, cubic, B-spline) were employed for classification.
• A soft-voting ensemble approach and five-fold cross-validation were utilized for model evaluation on the Lung-PET-CT-DX dataset.

Main Results:

• The proposed LC detection model achieved a high accuracy of 99.0%.
• The model demonstrated a minimal loss of 0.07, indicating strong performance.
• The classification process required limited computational resources.

Conclusions:

• The developed model shows significant potential for accurate and optimal LC detection using PET/CT scans.
• The findings can enhance clinical practice by providing sophisticated and interpretable diagnostic outcomes.
• Future work can explore advanced feature fusion techniques to further improve the model's capabilities.