The Surgical Planning Challenge Pulmonary ground-glass nodules (GGNs) span a spectrum of invasiveness - from benign atypical adenomatous hyperplasia (AAH) to fully invasive adenocarcinoma (IA). The surgical approach depends critically on this invasiveness: lobectomy is standard for invasive disease, while sublobectomy suffices for minimally invasive or non-invasive lesions.
The Mismatch Problem Without accurate preoperative staging, surgeons may perform lobectomies for non-invasive GGNs (overtreatment) or sublobectomies for invasive tumors (undertreatment). Both mismatches have clinical consequences: unnecessary parenchyma loss reduces pulmonary function, while inadequate resection increases recurrence risk.
The Novel Approach This multicenter study developed a framework integrating CT-based radiomics with 3D deep learning to predict GGN infiltration status preoperatively. The system combines handcrafted radiomic features (shape, texture, intensity statistics) with automatically learned deep learning features - fusing complementary strengths of both methodologies.
Multicenter Validation Design 354 patients from Changzhou First People's Hospital served as the test set, with external validation on 54 patients from Zhongshan Hospital (Fudan University) and 32 patients from Changzhou Second People's Hospital. Multi-institution validation is essential given known scanner-dependent variability in CT radiomic features.
3D ROI Segmentation Tumor regions of interest (ROIs) were manually delineated on thin-slice CT images (less than 2mm slice thickness) using ITK-SNAP software. Two surgeons independently performed 3D ROI mapping with consensus reconciliation, minimizing inter-observer segmentation variability. Only solitary GGNs under 2cm with solid component proportion below 50% were included.
Multi-Category Feature Extraction Features were extracted from four categories: morphological descriptors (shape, volume, surface area), first-order statistical metrics (intensity histogram statistics), texture patterns (GLCM, GLRLM features capturing spatial relationships between pixels), and higher-order wavelet transform features. This comprehensive extraction produced a large initial feature pool.
3D CNN with Data Augmentation Deep learning features were extracted using a 3D convolutional neural network with 13 convolutional layers and 3 fully connected layers. Pre-trained weight parameters were loaded to initialize the network, and random rotations at varying angles were used for data augmentation, improving model robustness to nodule orientation variability.
LASSO Feature Selection and SVM Classification Clinical, radiomics, and deep learning features were integrated into a unified feature vector. LASSO regression with k-fold cross-validation selected the most discriminative invasion-associated features. A nonlinear SVM with radial basis function kernel then performed binary classification (invasive vs. non-invasive). SMOTE oversampling balanced the training set.
AUC Performance The integrated radiomics-neural network model achieved an AUC of 0.85 on the primary test set. External validation sets achieved AUCs of 0.66 and 0.71, demonstrating reasonable generalizability across institutions with different CT scanners and patient populations - though the drop from training to external validation reflects typical overfitting in radiomic models.
Lobectomy-Sublobectomy Mismatch Reduction The model's clinical impact was quantified by its reduction in surgical mismatch: the predicted mismatch rate between lobectomy and sublobectomy recommendations was 21.48%, representing a 35.57% reduction compared to standard clinical approach without the model. This is a substantial improvement in surgical decision accuracy.
Within-Sublobectomy Mismatch Reduction Even within the sublobectomy group, the model reduced mismatches (between different sublobectomy extents) by 13.66%, reaching an error rate of 10.73%. This shows the model provides value not just at the lobectomy vs. sublobectomy decision boundary but also in guiding the extent of limited resection.
Integration Superiority The combined radiomics + neural network model outperformed both pure radiomics and pure CNN models, and also outperformed SVM-only and traditional radiomics approaches. This confirms that feature fusion of complementary handcrafted and learned features provides additive value over any single approach.
Avoiding Overtreatment For non-invasive GGNs (AIS, AAH), the model's prediction could guide surgeons toward sublobectomy or even watchful waiting, preserving lung tissue and pulmonary function. Given that GGN patients often have excellent prognoses, avoiding unnecessary lobectomy has significant quality of life implications.
Ensuring Adequate Resection Conversely, for GGNs predicted to be invasive adenocarcinoma, the model justifies more extensive resection with systematic lymph node dissection. Under-resecting invasive GGNs risks positive margins and local recurrence - both oncologically consequential.
Integration with CT Screening Programs As LDCT lung cancer screening expands in China and globally, the volume of incidental GGNs requiring management decisions is growing rapidly. An automated GGN invasion classifier could triage CT findings, flagging which nodules require surgical consultation versus continued surveillance.
Reducing Need for Intraoperative Frozen Section Currently, frozen section pathology during surgery is used to determine invasion status and thus the appropriate surgical extent. If preoperative AI prediction is sufficiently accurate, it could reduce reliance on frozen section, simplifying intraoperative workflow and potentially shortening operative time.
AUC Drop in External Validation The external validation AUCs of 0.66 and 0.71 (vs. 0.85 in the test set) suggest meaningful performance degradation across institutions. This likely reflects scanner-dependent radiomic feature instability and the relatively small external validation cohorts. Larger multicenter prospective studies are needed.
Manual Segmentation Bottleneck ITK-SNAP manual segmentation is time-consuming and requires trained personnel. For clinical deployment at scale, automated GGN segmentation algorithms must be integrated upstream of the feature extraction pipeline.
Pure GGN Inclusion Criteria The study included only solitary GGNs under 2cm with solid component under 50%. Multiple GGNs, larger lesions, and lesions with greater solid components are excluded, limiting applicability to a subset of clinical GGN scenarios.
Future Directions Priority next steps include developing a fully automated pipeline from CT to invasiveness prediction without manual segmentation, testing the model on prospectively collected GGN cohorts, extending to larger and multiple GGNs, and eventually integrating the prediction into thoracic surgery decision support systems used at the point of care.