200 - Radiological prior Knowledge-Guided Deep Learning for Predicting Invasiveness of Medium-Sized Pulmonary Nodules

Purpose/Objective(s):

Accurate noninvasive assessment of pulmonary nodule invasiveness is critical for clinical decision-making, particularly selecting management strategies such as surgical resection or stereotactic body radiotherapy. However, for medium-sized nodules (10-20 mm), subtle morphologic differences and substantial inter-reader variability limit the reliability of both conventional radiological assessment and existing deep learning approaches. To address this gap, we developed a radiological prior knowledge-enhanced deep learning framework that improves preoperative invasiveness prediction from chest CT, thereby supporting risk-adapted thoracic oncology management.

Materials/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 for generalizability evaluation. The proposed dual-scale, multi-task deep learning framework (DSMT-Net) adopts a dual-encoder architecture to capture both local nodule morphology and global pulmonary context. To incorporate radiological prior knowledge, the framework augments the primary invasiveness prediction by concurrently predicting bubble lucency and pleural retraction, two CT findings associated with invasive pathology, thereby strengthening feature learning. An auxiliary segmentation branch was also integrated to improve spatial awareness and feature alignment. A two-stage training strategy was employed, consisting of joint multi-task optimization followed by invasiveness-specific fine-tuning. Model performance was assessed using the area under the curve (AUC), accuracy, 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 an AUC of 0.902, outperforming the baseline single-task model (AUC=0.851). DSMT-Net also improved diagnostic accuracy compared to manual radiologist assessment (82.76% vs. 72.41%). On the independent external cohort, DSMT-Net maintained robust performance with an AUC of 0.864 and an accuracy of 77.56%, despite substantial variability in acquisition protocols and imaging quality.

Conclusion:

DSMT-Net improves noninvasive prediction of invasiveness in medium-sized pulmonary nodules by integrating dual-scale volumetric features with radiological prior knowledge. The framework outperforms baseline deep learning models and experienced radiologists, demonstrating strong generalizability. This approach has high potential as a clinical decision-support tool for personalized treatment stratification and planning in thoracic oncology, including guidance on surgery versus radiotherapy selection.