Predicting Lung Nodule Growth from Follow-up CT Scans with Deep Isotropic Network

Lung nodule growth prediction is vital for improving lung cancer diagnosis. The irregular time intervals in clinical follow-up 3D CT scans and the complex nature of lung nodules present significant challenges for accurate modeling nodule growth. Leveraging these clinical follow-up data is crucial for enhancing prediction models. In this study, we propose Lung Nodule Growth Network (LNGNet), a two-stage novel model based on our self-constructed temporal CT dataset named LNt. LNGNet features an isotropic-MedNeXt backbone for superior feature extraction, the Adaptive Temporal Scaling (ATS) Module to handle irregular time intervals, and an autoencoder with attention feature maps to guide texture generation at future time points. LNGNet accurately predicts nodule mass and volume and generates high-quality shape and texture visualizations, with 5% improvement on volume prediction AUC, 7.2% improvement on mass prediction AUC and 2.89% improvement on dice score. These advancements demonstrate LNGNet’s potential for clinical application, enhancing early diagnosis and treatment strategies for lung cancer patients.