Dsmt-Net: A Dual-Scale Multi-Task Learning Framework for Invasiveness Prediction of Medium-Sized Pulmonary Nodules
Abstract
Purpose
Non-invasive diagnosis of pulmonary nodule invasiveness via CT imaging is critical for clinical management. However, assessing the status of medium-sized nodules (10-20 mm) remains a significant challenge for both radiologists and deep learning models. To address this, we developed DSMT-Net, a dual-scale, multi-task framework guided by radiologic prior knowledge, specifically designed to enhance prediction accuracy for these challenging nodules to support surgical decision-making.
Methods
This retrospective study utilized a multi-source dataset with pathologically verified invasiveness labels. The internal cohort was partitioned into training (n = 387) and testing (n = 87) sets, while an independent external cohort (n = 274) was used to evaluate generalizability. DSMT-Net employs a dual-encoder architecture to simultaneously capture local morphological details and global pulmonary context. The model integrates an auxiliary segmentation branch for spatial awareness and three parallel classification tasks: invasiveness prediction, bubble lucency detection, and pleural retraction identification, effectively incorporating expert radiologic prior knowledge. A two-stage training strategy was implemented: joint multi-task optimization followed by invasiveness-specific fine-tuning. Model performance was evaluated using the area under the curve(AUC), sensitivity, specificity, and F1-score. Furthermore, the performance of DSMT-Net was directly compared against that of senior radiologists.
Results
On the internal test set, DSMT-Net achieved a significantly higher AUC (0.902) than the baseline single-task model (AUC: 0.851). Additionally, DSMT-Net demonstrated superior diagnostic accuracy (82.76%) compared to manual radiological assessment (72.41%). Crucially, even across heterogeneous imaging protocols, DSMT-Net exhibited strong generalizability on the external validation set, with an accuracy of 77.56% and AUC of 0.864.
Conclusion
DSMT-Net significantly outperforms baseline models and human experts in assessing the invasiveness of medium-sized pulmonary nodules. By effectively fusing dual-scale volumetric features with auxiliary radiological signs, this framework provides a high-precision, automated solution for preoperative surgical planning, demonstrating strong potential for clinical adoption.