This paper proposes a novel dynamic difference-aware temporal residual network (DDaTR) for longitudinal radiology report generation (LRRG). Unlike existing LRRG methods that simply extract features from previous and current images and concatenate them, DDaTR introduces two modules—the dynamic feature alignment module (DFAM) and the dynamic difference recognition module (DDAM)—to capture multi-level spatial correlations. DFAM aligns previous features across multiple images, and DDAM effectively captures inter-examination difference information based on this alignment. Furthermore, it effectively models temporal correlations using the dynamic residual network. Experimental results demonstrate that our proposed method outperforms existing methods across three benchmarks.
