Radiology report generation (RRG) is a demanding yet challenging task that involves producing multi-sentence diagnostic narratives, requiring long-form text with high diversity while addressing inherent data bias. Sentence-level diversity is therefore crucial for capturing varying diagnostic details across multiple regions of interest
(ROIs) within a single report, yet it remains underexplored in the field. In this paper, we propose DADNET, a novel diffusion-based framework that leverages the inherent ability of diffusion models to generate diverse text. We make the first attempt to integrate large language models (LLMs) to bridge the inherent training-inference gap in diffusion models. Specifically, LLMs are used to draft a preliminary report, which is subsequently incorporated into the diffusion process to enhance report diversity. Additionally, we introduce a bias equalization technique using domain-specific priors to mitigate data distribution biases, improving the quality and reliability of generated reports under various scenarios. Experimental results
demonstrate that DADNET outperforms existing approaches under the same non-autoregressive (NAR) mechanism and sets a new benchmark for diversity in RRG. This work underscores the importance of diversity in RRG and establishes DADNET as a pioneering framework for addressing this challenge with NAR methods.
