Related Experiment Video
Updated: Jun 10, 2025
10:18
Dynamic Pore-scale Reservoir-condition Imaging of Reaction in Carbonates Using Synchrotron Fast Tomography
Published on: February 21, 2017
8.5K
Echo from noise: synthetic ultrasound image generation using diffusion models for real image segmentation.
David Stojanovski1, Uxio Hermida1, Pablo Lamata1
1King's College London, School of Biomedical Engineering & Imaging Sciences, London, SE1 7EU, UK.
Simplifying Medical Ultrasound : 4Th International Workshop, ASMUS 2023, Held in Conjunction with MICCAI 2023, Vancouver, BC, Canada, October 8, 2023, Proceedings. ASMUS (Workshop) (4Th : 2023 : Vancouver, B.C. ; Online)
|October 15, 2024
Summary
We developed a new method using Denoising Diffusion Probabilistic Models (DDPMs) to create synthetic ultrasound images for training AI models. These AI models show improved cardiac segmentation performance on real images.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Deep learning models require large datasets for training.
- Acquiring diverse and annotated medical imaging data, such as ultrasound images, can be challenging and time-consuming.
- Synthetic data generation offers a potential solution to data scarcity in medical AI.
Purpose of the Study:
- To introduce a novel pipeline for generating synthetic ultrasound images using Denoising Diffusion Probabilistic Models (DDPMs).
- To evaluate the efficacy of these synthetic images as a substitute for real data in training deep learning models for cardiac image analysis.
- To demonstrate improved performance in cardiac segmentation tasks using AI models trained on synthetic data.
Main Methods:
- Utilized Denoising Diffusion Probabilistic Models (DDPMs) guided by cardiac semantic label maps to generate synthetic ultrasound images.
- Generated synthetic 2D echocardiograms.
- Trained a neural network for left ventricle and left atrium segmentation exclusively on the generated synthetic images.
Main Results:
- The neural network trained on synthetic images achieved high Dice scores on an unseen real dataset: 88.6% (left ventricular endocardium), 91.9% (left ventricular epicardium), and 85.2% (left atrium).
- Demonstrated significant relative improvements in Dice scores compared to the previous state-of-the-art: 9.2% (LV endocardium), 3.3% (LV epicardium), and 13.9% (LA).
- Validated the potential of synthetic ultrasound images for training robust deep learning models.
Conclusions:
- The proposed DDPM-guided pipeline effectively generates high-quality synthetic ultrasound images.
- Synthetic ultrasound images can successfully replace real data for training deep learning models in cardiac segmentation.
- This approach offers a promising solution for data augmentation and can be extended to other medical imaging modalities and tasks.
