CDPM-Align: Multi-Scale Guidance-Aligned Diffusion Pretraining

This is the academic project page for the MICCAI 2026 accepted paper:

CDPM-Align: Multi-Scale Guidance-Aligned Diffusion Pretraining for Robust Few-Shot Anatomical Landmark Detection

Paper Details

• Venue: MICCAI 2026 (Accepted)
• Authors: Roberto Di Via, Irina Voiculescu, Francesca Odone, Vito Paolo Pastore
• Affiliations: MaLGa-DIBRIS, University of Genoa | Oxford University

Links

• Code (GitHub)

Abstract

Anatomical landmark detection is a fundamental task in medical image analysis supporting a wide range of diagnostic and interventional workflows. Although recent methods have achieved sub-millimetric localisation, accuracy alone is not sufficient for clinical deployment, requiring reliability and robustness in prediction.

In this work, we introduce CDPM-align, a multi-scale guidance-aligned conditional diffusion pre-training for anatomical landmark detection. Our experimental setup focuses on few-shot regimes (10 and 25 annotated images). Specifically, we employ three popular heterogeneous small-scale benchmark datasets for representation learning via conditional generative pre-training.

Our results confirm that generative pre-training enables the model to learn robust representations. This improves both accuracy and uncertainty on the downstream tasks, advancing towards safe and efficient clinical deployment.

Key Results

• State-of-the-art few-shot performance: Best MRE in 5/6 dataset-budget settings
• Sub-millimetre ERE: At 25-shot, achieves ERE below 1mm on ISBI2015 (0.95mm) and DHA (0.97mm)
• +22% improvement on ISBI2015 10-shot (2.11 vs 2.70 mm)
• +42% improvement on DHA 10-shot (2.51 vs 4.34 mm)
• Data-efficient: Pretrained on only 988 images, matches or surpasses models trained on 112k NIH radiographs

Method

We propose CDPM-align, a conditional diffusion pretraining framework with multi-scale guidance alignment. The method consists of:

1. Conditional Diffusion Probabilistic Model (CDPM): Training a class-conditional diffusion model with classifier-free guidance on small-scale heterogeneous X-ray datasets

2. Multi-scale Alignment: Enforcing directional consistency of the guidance signal across independently sampled diffusion timesteps and UNet hierarchy levels

3. Downstream Fine-tuning: End-to-end fine-tuning of the pretrained backbone for anatomical landmark detection with pixel-wise classification

Datasets

• Shenzhen: 279 chest radiographs, 6 landmarks
• ISBI2015: 400 lateral cephalograms, 19 landmarks
• Digital Hand Atlas (DHA): 910 hand radiographs, 37 landmarks

Website

This page is built using the Academic Project Page Template and is hosted on GitHub Pages.

License

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.