Graph-Driven Medical Report Generation with Adaptive Knowledge Distillation

Abstract

Automated medical report generation (MRG) faces a critical hurdle in seamlessly integrating detailed visual evidence with accurate clinical diagnoses. Current approaches often rely on static knowledge transfer, overlooking the complex interdependencies among pathological findings and their nuanced alignment with visual evidence, often yielding reports that are linguistically sound but clinically misaligned. To address these limitations, we propose a novel graph-driven medical report generation framework with adaptive knowledge distillation. Our architecture leverages a dual-phase optimization process. First, visual–semantic enhancement proceeds through the explicit correlation of image features with a structured knowledge network and their concurrent enrichment via cross-modal semantic fusion, ensuring that generated descriptions are grounded in anatomical and pathological context. Second, a knowledge distillation mechanism iteratively refines both global narrative flow and local descriptive precision, enhancing the consistency between images and text. Comprehensive experiments on the MIMIC-CXR and IU X-Ray datasets demonstrate the effectiveness of our approach, which achieves state-of-the-art performance in clinical efficacy metrics across both datasets.