NGP-Net: A Lightweight Growth Prediction Network for Pulmonary Nodules

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Radiology research

Background:

• Pulmonary nodule growth monitoring is vital for lung cancer prevention but faces challenges with subtle growth patterns and irregular CT scan intervals.
• Current methods often rely on single-timepoint analyses or fixed intervals, limiting predictive accuracy in dynamic clinical scenarios.
• Accurate prediction of nodule growth is essential for early detection and intervention in lung cancer management.

Purpose of the Study:

• To introduce NGP-Net, a novel W-shaped architecture for dynamic pulmonary nodule growth prediction using irregularly sampled longitudinal CT scans.
• To develop a model capable of learning temporal dynamics from sparse data and reconstructing nodule features at future timepoints.
• To enhance the accuracy and reliability of pulmonary nodule growth assessment in clinical practice.

Main Methods:

• Proposed NGP-Net, a W-shaped deep learning architecture featuring a Spatial-Temporal Encoding Module (STEM) for irregular data.
• Developed a dual-branch decoder for high-fidelity reconstruction of nodule textures and shapes at arbitrary future timepoints.
• Curated and released a dataset of 378 chest CT scans with 226 pulmonary nodules from 103 patients, featuring longitudinal follow-up (2-64 months) and radiologist annotations.

Main Results:

• NGP-Net achieved state-of-the-art performance on the newly curated dataset, demonstrating superior predictive accuracy.
• Achieved the lowest mean square error (6.13 × 10^-3 overall, 1.28 × 10^-4 nodule-specific) and significant improvements in Dice similarity coefficient (10.55%), PSNR (0.29 dB), and SSIM (5.94%).
• The model showed robust and precise predictions across diverse nodule growth scenarios, validating its clinical utility.

Conclusions:

• NGP-Net effectively addresses the limitations of existing methods for pulmonary nodule growth prediction from irregular CT scans.
• The proposed architecture and dataset provide a valuable tool for advancing lung cancer monitoring and early detection.
• NGP-Net's performance indicates its potential to significantly support radiologists in clinical decision-making for pulmonary nodule management.