Clinically Grounded Curriculum Learning for Radiology Report Generation

Current radiology report generation (RRG) systems predominantly rely on random mini-batch sampling, a paradigm that treats all reports equally and ignores the pathological complexity across different patients. Consequently, while producing fluent text, these models struggle to accurately capture fine-grained local pathology attributes, often suffering from severe clinical hallucinations. Curriculum learning presents a promising solution by progressively introducing data from easy to hard; however, defining a valid difficulty metric in the RRG context is non-trivial, as naive proxies like report length cannot effectively reflect true pathological complexity. To address this, we propose a Clinically-structured Curriculum RRG framework (C²-GEN). We instantiate this framework with hierarchical difficulty metrics that assess clinical complexity through both the global semantic density of pathology findings and the structural medical knowledge. Guided by these metrics, our approach replaces conventional uniform sampling with a difficulty-aware training schedule, gradually expanding the training pool towards complex cases while continually replaying simpler anchor samples to prevent catastrophic forgetting. Furthermore, we augment the curriculum from the aspect of network architecture, introducing an adaptive memory mechanism that dynamically adjusts the attention mechanism based on sample complexity to better bridge visual features and complex text generation. Experiments on public datasets demonstrate the effectiveness of C²-GEN, consistently improving the fidelity of local pathology attributes compared to existing baselines. Comprehensive analyses illustrate the progressive learning patterns and cross-modal alignment in RRG, where these insights establish curriculum learning as a principled pathway to mitigating clinical hallucinations.