Explainable Vision-Language Framework for Automated Lung Nodule Risk Stratification Using Dual-Attention Segmentation and Large Medical Models

Abstract

The clinical management of pulmonary nodules detected in low‑dose computed tomography scans relies critically on accurate risk stratification to distinguish benign from malignant lesions while minimizing unnecessary invasive procedures. Existing deep learning approaches often operate as opaque classifiers, offering little insight into the visual and semantic rationale behind their predictions. This paper introduces an explainable vision‑language framework that integrates a dual‑attention segmentation backbone with large medical vision‑language models to automate lung nodule risk assessment. The proposed architecture first isolates nodule regions through a path‑aggregation encoder combined with channel‑wise and spatial attention mechanisms, producing high‑fidelity segmentation masks that are subsequently analyzed by a multimodal transformer that encodes both radiological features and structured clinical text. A dedicated explainability module generates natural‑language justifications aligned with segmented regions, thereby enabling clinicians to inspect the decision‑making process at both pixel and concept levels. The paper discusses structural trade‑offs between segmentation fidelity, model interpretability, and computational efficiency, and examines deployment considerations including data governance, infrastructure scalability, and regulatory compliance. Fairness and robustness are analyzed across demographic subgroups and imaging acquisition protocols, and policy implications for integrating such systems into existing radiology workflows are explored. By bridging the gap between high‑accuracy black‑box models and the demand for transparent reasoning in high‑stakes medical decisions, the proposed framework advances the state of the art in trustworthy AI for thoracic oncology.