AnchorDiff: Topology-Aware Masked Diffusion with Confidence-based Rewriting for Radiology Report Generation
The paper introduces AnchorDiff, a novel framework for generating radiology reports using a topology-aware masked diffusion approach. This method addresses limitations of traditional autoregressive models by incorporating bidirectional context and clinical anchors. Extensive experiments demonstrate that AnchorDiff achieves state-of-the-art performance in radiology report generation.
- ▪AnchorDiff integrates knowledge-graph-derived clinical anchors into diffusion language modeling.
- ▪The framework employs a topology-aware training strategy to enhance token generation.
- ▪Extensive experiments show that AnchorDiff outperforms existing methods in radiology report generation.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Artificial Intelligence arXiv:2605.17071 (cs) [Submitted on 16 May 2026] Title:AnchorDiff: Topology-Aware Masked Diffusion with Confidence-based Rewriting for Radiology Report Generation Authors:Shiying Yu, Jielei Wang, Guoming Lu View a PDF of the paper titled AnchorDiff: Topology-Aware Masked Diffusion with Confidence-based Rewriting for Radiology Report Generation, by Shiying Yu and 2 other authors View PDF HTML (experimental) Abstract:Radiology report generation (RRG) aims to automatically produce clinically accurate textual reports from medical images. Existing methods predominantly rely on autoregressive (AR) language models, whose causal dependency structure restricts generation to a unidirectional left-to-right process.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
