Convergent Generative Adversarial Networks for Medical Report Generation With Visual Manifold Graph
Medical report generation aims to reduce the reporting burden of radiologists by translating medical images into clinically meaningful descriptions. Although generative adversarial networks (GANs) have shown potential for image captioning and report generation, their application to radiology reports remains challenging because radiology text requires standardized negative expressions, adversarial training may be unstable, and lexical metrics alone cannot fully reflect clinical utility. To address these issues, we propose ConCapGAN, a GAN-based medical report generation framework with a visual manifold graph (VMG), a zero-centered Wasserstein regularizer, and a language-style constraint. The VMG explicitly models object–predicate associations between detected image regions and report semantics, thereby helping represent clinically important positive and negative findings. The zero-centered regularizer improves the local stability of adversarial optimization under clearly stated assumptions, and the language-style constraint encourages reports to follow clinician-like phrasing without introducing an additional large language module. Experiments on IU X-Ray, MIMIC-CXR, and LGK show that ConCapGAN achieves competitive natural language generation performance and improved clinical efficacy (as measured by several diagnostic metrics) compared with recent report generation methods.