Voxel Normalization in LDCT Imaging: Its Significance in Texture Feature Selection for Pulmonary Nodule Malignancy...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Lung cancer is a leading global cause of cancer mortality.
• Early detection using low-dose computed tomography (LDCT) is challenged by a lack of objective criteria and manual interpretation needs.
• Public datasets often lack pathology confirmation and have variable voxel sizes, impacting model reliability.

Purpose of the Study:

• To assess the impact of voxel normalization on pulmonary nodule classification.
• To introduce a Fast Fourier Transform (FFT)-based contour fusion method for interpretable nodule annotation integration.
• To develop and evaluate machine learning and deep learning models for nodule classification using processed LDCT data.

Main Methods:

• Voxel normalization was applied to LDCT images from 415 patients across two centers.
• Machine learning and deep learning models, including transformers with attention mechanisms, were developed.
• A novel FFT-based contour fusion method was introduced as an interpretable alternative to generative adversarial networks.

Main Results:

• Voxel normalization improved feature overlap between datasets by 64%, enhancing selection stability.
• Machine learning models achieved 92.6% accuracy, while deep learning models reached 98.5% accuracy.
• The FFT-based method provided clinically interpretable integration of expert annotations, overcoming GAN limitations.

Conclusions:

• Voxel normalization enhances the reliability of pulmonary nodule classification.
• The FFT-based method presents a viable approach for interpretability in deep learning applications for medical imaging.
• Further multi-center research is recommended to explore the implications of these findings.