Core Concepts
XReal introduces a novel controllable diffusion model for generating realistic chest X-ray images with precise anatomy and pathology control, outperforming existing models in medical imaging applications.
Abstract
Abstract:
Large-scale generative models struggle with image hallucination and anatomically inaccurate outputs due to reliance on textual inputs.
XReal presents a controllable diffusion model for realistic chest X-ray generation with spatial control over anatomy and pathology.
Introduction:
Deep generative models excel in various applications but face challenges in medical imaging.
Text-to-image models lack spatial control over critical information in medical images.
Method:
XReal consists of an Anatomy Controller, Latent Diffusion Model, and Pathology Controller for precise image generation.
The Anatomy Controller controls anatomical structure, while the Pathology Controller infuses pathology into the generated image.
Experiments:
Trained on the MIMIC-CXR dataset, XReal outperforms existing models in quantitative metrics and expert radiologist evaluation.
Offers high-quality, clinically realistic X-ray images with precise anatomy and pathology manifestation.
Results:
XReal demonstrates superior performance compared to state-of-the-art methods in anatomical realism and pathology control.
Expert radiologist evaluation confirms the clinical realism of generated images.
Discussion:
XReal's lightweight design enables precise anatomy and pathology control, enhancing its utility in medical imaging applications.
Image editing capabilities allow for counterfactual generation, disease simulation, and training purposes.
Conclusion:
XReal is the first diffusion model offering precise anatomy and pathology control for realistic chest X-ray generation.
Plans to extend the work to other modalities like MRI and CT scans are mentioned.
Stats
XRealは、専門家の放射線科医の評価に基づいて、解剖学と病理学の現実性を示す13%および10%のゲインを示しています。