Abstract
Diffusion models have shown strong potential for medical image generation, yet their clinical applicability critically depends on the anatomical consistency of generated region-of-interest (ROI) areas, which are essential for supporting more accurate diagnosis. In real-world settings, ROI annotations are often sparse and time-consuming, resulting in a partial-label problem. Although Classifier-Free Guidance (CFG) enhances conditional control, a global CFG scale applied uniformly across the image space can misguide unannotated regions, leading to spatial inconsistencies and degraded clinical reliability. In this work, we present VGR-Diff, a validity-gated ROI guidance latent diffusion framework that robustly leverages sparse ROI annotations for medical image generation. Our method integrates three key components: 2.5D biased slice sampling to enhance cross-slice contextual consistency, validity-gated ROI supervision during training to prevent erroneous guidance from missing labels, and ROI-aware CFG during inference to decouple global anatomical structure from local ROI constraints. Extensive experiments on a public brain MRI dataset (BraTS2023) and a private pelvic MRI–CT dataset from Qilu hospital demonstrate that, under various annotation sparsity conditions, VGR-Diff consistently achieves superior ROI fidelity among state-of-the-art 2D and 3D generative models, while maintaining competitive global image quality. By minimizing dependence on dense annotations and ensuring anatomically consistent generation, VGR-Diff offers a practical and clinically viable solution for real-world diffusionbased medical image generation.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/DGM4MICCAI_007.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/profile?id=~Chang_Li1
BibTex
@InProceedings{LiCha_Making_MICCAISAT2026,
author = { Li, Chang AND Zhao, Xiang AND Moraros, John AND Meng, Jia AND Liu, Jingxin AND Wang, Kexin AND Wang, Shuihua},
title = { { Making sparse labels reliable: validity-gated ROI guidance for medical image generation with conditional latent diffusion model } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17258},
month = {pending},
page = {pending}
}