.png){kind=icon}
.png){kind=icon}
Return
UniMedDiff: a knowledge-enhanced diffusion model for medical image generation from clinical reports
.png){kind=icon}
OA
AI
L
P
B
S
J
Z
Y
Y
DOI：10.1038/s41746-026-03135-x.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The scarcity of high-quality medical data restricts the advancement of computer-aided diagnosis. Text-to-image (TTI) generation offers a promising solution through controlled and diverse synthesis but is still hindered by noisy data, complex report structures, and varied disease representations. We propose UniMedDiff, a disease-knowledge-enhanced diffusion framework for synthesizing anatomically accurate and pathologically diverse medical images from clinical reports. Trained on noise-filtered image–text pairs, UniMedDiff extracts and aligns concise text embeddings with visual features from lengthy reports, enabling pathology-aware guidance for medical image generation. It facilitates controllable generation by integrating reports and prior disease knowledge into the diffusion process, achieving favorable performance across multiple datasets and evaluation tasks. On chest X-rays (CXRs), UniMedDiff generates diverse CXRs representing 11 pulmonary pathologies from both in-/out-of-distribution reports. Qualitative and quantitative results demonstrate its state-of-the-art performance in generation quality, computational efficiency and factual correctness. UniMedDiff also exhibits utility in downstream tasks, including unimodal classification and cross-modal report generation, achieving near full-data performance by augmenting 1% real data. Its cross-modality generalization is evaluated through dermoscopy and colonoscopy image synthesis. By bridging clinical reports and visual disease patterns, UniMedDiff advances medical image synthesis for diagnostic research and data augmentation.
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.