Scholarlink

Latent Diffusion models (LDMs) require large annotated datasets for training, limiting their applicability in medical imaging where datasets are typically smaller and sparsely annotated. We introduce DiNO-Diffusion, a self-supervised method for training LDMs that conditions the generation process on image embeddings extracted from DiNO, a pretrained vision transformer. By not relying on annotations, our training leverages over 868k unlabelled images from public chest X-Ray (CXR) datasets. DiNO-Diffusion shows comprehensive manifold coverage, with FID scores as low as 4.7, and emerging properties when evaluated in downstream tasks, allowing to generate semantically-diverse synthetic datasets even from small data pools, demonstrating up to 20% AUC increase in classification performance when used for data augmentation. Results suggest that DiNO-Diffusion could facilitate the creation of large datasets for flexible training of downstream AI models from limited amount of real data, while also holding potential for privacy preservation. Additionally, DiNO-Diffusion demonstrates zero-shot segmentation performance of up to 84.4% Dice score when evaluating lung lobe segmentation, evidencing good CXR image-anatomy alignment akin to textual descriptors on text-to-image LDMs. Finally, DiNO-Diffusion can be easily adapted to other medical imaging modalities or state-of-the-art diffusion models (DMs), allowing large-scale, multi-domain image generation pipelines for medical imaging.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DiNO-Diffusion: Scaling Medical Diffusion Models via Self-Supervised Pre-Training

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Guillermo Jimenez-Perez,
  • Pedro Osório,
  • Josef Cersovsky,
  • Javier Montalt-Tordera,
  • Jens Hooge,
  • Steffen Vogler,
  • Sadegh Mohammadi

摘要

Latent Diffusion models (LDMs) require large annotated datasets for training, limiting their applicability in medical imaging where datasets are typically smaller and sparsely annotated. We introduce DiNO-Diffusion, a self-supervised method for training LDMs that conditions the generation process on image embeddings extracted from DiNO, a pretrained vision transformer. By not relying on annotations, our training leverages over 868k unlabelled images from public chest X-Ray (CXR) datasets. DiNO-Diffusion shows comprehensive manifold coverage, with FID scores as low as 4.7, and emerging properties when evaluated in downstream tasks, allowing to generate semantically-diverse synthetic datasets even from small data pools, demonstrating up to 20% AUC increase in classification performance when used for data augmentation. Results suggest that DiNO-Diffusion could facilitate the creation of large datasets for flexible training of downstream AI models from limited amount of real data, while also holding potential for privacy preservation. Additionally, DiNO-Diffusion demonstrates zero-shot segmentation performance of up to 84.4% Dice score when evaluating lung lobe segmentation, evidencing good CXR image-anatomy alignment akin to textual descriptors on text-to-image LDMs. Finally, DiNO-Diffusion can be easily adapted to other medical imaging modalities or state-of-the-art diffusion models (DMs), allowing large-scale, multi-domain image generation pipelines for medical imaging.