Tumor Semantic Modeling with Vascular Morphology Prior Correction Reduces False-Positive Malignant Predictions on Contrast-Enhanced Lung CT.
Authors
Affiliations (5)
Affiliations (5)
- Department of Electrical Engineering, Shenyang University of Technology, Shenyang, Liaoning, China.
- Department of Radiation Oncology, General Hospital of Northern Theater Command, Shenyang, Liaoning, China.
- Dalian Women and Children's Medical Center (Group), Dalian, Liaoning, China.
- Department of Radiation Oncology, General Hospital of Northern Theater Command, Shenyang, Liaoning, China. [email protected].
- Department of Electrical Engineering, Shenyang University of Technology, Shenyang, Liaoning, China. [email protected].
Abstract
To develop an interpretable contrast-enhanced CT-based framework for benign-malignant lung tumor classification by integrating radiology-style semantic representations with global pulmonary vascular morphology features. Two particular difficulties with this task lie in (1) establishing an interpretable model to distinguish tumor data that are difficult to differentiate between benign and malignant in clinical practice, and (2) based on clinical experience, incorporate the vascular characteristics around the tumor into the decision-making process. We propose a vascular knowledge-enhanced classification framework that integrates semantic tumor representations extracted by large multimodal models with peritumoral vascular morphological features. The framework consists of three components: (1) an interpretable feature extractor for tumor representation learning, (2) vascular morphology modeling derived from vessel segmentation, and (3) a feature fusion module combining tumor and vascular information for classification. In a retrospective cohort of 420 patients with pathologically confirmed pulmonary tumors, the proposed method was assessed using patient-level, stratified five-fold cross-validation. Relative to the semantic-only branch, decision-level fusion increased accuracy (71.4% vs. 75.5%), AUC (0.764 vs. 0.816), and specificity (57.3% vs. 77.8%), but decreased sensitivity (81.1% vs. 73.9%). These findings indicate that vascular prior-guided correction primarily enhanced benign-case recognition and suppressed false-positive malignant predictions, at the expense of sensitivity for malignant cases. In single-center validation, pulmonary vascular morphology offered complementary structural evidence for differentiating benign from malignant pulmonary tumors larger than 3 cm. The proposed AND-like decision-level correction rule enhanced specificity by reducing false-positive malignant predictions.