Comorbidity-Aware Radiology Report Generation

Chest X-ray report generation systems are valuable for assisting disease diagnosis and improving healthcare efficiency. However, existing methods still face two key challenges. First, multiple diseases often co-occur, leading to a combinatorial explosion of label combinations and sparse supervision for learning a generalizable image-to-label mapping. Second, standard autoregressive training is primarily optimized for local linguistic fluency, but provides limited direct supervision for global clinical fidelity, often resulting in omissions or contradictions in complex multi-abnormality cases. To address these issues, we introduce a geometry-aware multimodal large-model framework. The framework leverages hyperbolic geometry as a hierarchy-aware inductive bias and applies hierarchy-consistent constraints at three levels: representation, loss, and reinforcement signals. Specifically, we build a hyperbolic contrastive language-image pre-training (CLIP) model to align image-text semantics; estimate class prototypes from single-label cases and perform prototype-driven hyperbolic contrastive learning to enforce radius ordering for multi-label samples; and incorporate a learnable hyperbolic soft reward into the group relative policy optimization (GRPO) framework to directly optimize clinical accuracy. On the MIMIC-CXR benchmark, HyCoRG obtains BLEU-4 of 0.136 as well as clinical F1 of 0.515, indicating improved disease-level clinical consistency. Ablation studies confirm that both the framework and the geometric constraints contribute independently and synergistically. Our approach improves both linguistic quality and diagnostic fidelity, and offers a transferable hierarchy-aware paradigm for multi-label medical text generation. The code is publicly available at https://github.com/hongzezhu/HyCoRG.