Integrating Diagnostic Guidance and Historical Context for Radiology Report Generation

Radiology report generation (RRG) is a challenging task to automatically generate diagnostic reports for radiology images. Despite advancements in RRG, accurately generating diagnostic reports remains difficult due to the imbalanced disease distribution and lack of temporal information, leading to unstable performance and unreliable results. To address these challenges, we propose a progression-aware model that improves RRG by incorporating diagnostic guidance and historical context. We begin with a multimodal enhancement module to integrate a patient’s current and prior features, followed by temporal encoding to capture differences across time points. A disease progression modeling module then generates disease-aware diagnoses to inform the decoding process. Finally, our diagnosis-driven adaptive decoder fully leverages disease progression cues and diagnostic tokens to produce accurate radiology reports. Experimental results on the L-MIMIC dataset demonstrate the effectiveness of our method.