Multi-task learning for predicting pulmonary nodule growth and follow-up volume

Hypothesis: The primary objective of this study is to develop an end-to-end deep learning framework based on multi-task learning to predict pulmonary nodule growth by jointly modeling nodule segmentation and visual follow-up image synthesis. By decoupling nodule growth into deformation and texture evolution, the model aims to enhance predictive accuracy and clinical applicability through improved regional focus and deep supervision strategies. Methods: We present MT-NoGNet, a dual-task network for pulmonary nodule growth prediction via simultaneous deformation-texture modeling. The framework employs a shared encoder with two decoders: a spatial transformer for volume change estimation and a texture predictor with adaptive normalization. A cross-task attention mechanism enforces consistency between morphological expansion and internal density evolution. Results: Evaluated on longitudinal CT scans from 246 patients at Shanghai Chest Hospital, the framework achieved that the predicted peak signal to noise ratio (PSNR) was 44.30, structural similarity index (SSIM) was 0.7776, and dice similarity coefficient (DSC) was 0.7823. Conclusions: This study establishes that multi-task learning model of deformation-texture features significantly enhances pulmonary nodule growth prediction accuracy while providing radiologists with interpretable visualizations of progression patterns, demonstrating substantial potential for optimizing clinical surveillance protocols.