Towards Unified Medical Multimodal Understanding and Generation

Clinical AI systems are typically specialised for either understanding or generation, limiting their use in real workflows that require both. This thesis presents a unified approach to medical multimodal intelligence across three axes: data, modelling, and system design. It contributes large-scale benchmarks, clinically grounded vision-language representation learning, and instruction-driven medical image synthesis. These are integrated into a single system that bridges understanding and generation through unified data formatting, training strategy, and architecture. Experiments across medical multimodal benchmarks demonstrate effective interleaved multimodal reasoning and bidirectional transfer between tasks, advancing toward general-purpose, clinically compatible AI.