Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Ultrasound Image-to-Video Synthesis via Latent Dynamic Diffusion Models

Tingxiu Chen, Yilei Shi, Zixuan Zheng, Bingcong Yan, Jingliang Hu, Xiao Xiang Zhu, Lichao Mou

TL;DR

The paper tackles the lack of labeled ultrasound video data by proposing a latent dynamic diffusion model (LDDM) that converts static ultrasound images into plausible dynamic videos. This two-stage approach first encodes video dynamics into a low-dimensional latent space via a video autoencoder and then uses a conditional diffusion model conditioned on the seed image to generate realistic latent trajectories, which are decoded into videos. Evaluations on BUSV and BUSI show that including LDDM-generated videos improves downstream ultrasound video classification and that LDDM outperforms prior image-to-video methods on several quality metrics, while remaining computationally efficient. The method provides a practical data augmentation tool for medical video analysis, and the authors release code to enable adoption and further development.

Abstract

Ultrasound video classification enables automated diagnosis and has emerged as an important research area. However, publicly available ultrasound video datasets remain scarce, hindering progress in developing effective video classification models. We propose addressing this shortage by synthesizing plausible ultrasound videos from readily available, abundant ultrasound images. To this end, we introduce a latent dynamic diffusion model (LDDM) to efficiently translate static images to dynamic sequences with realistic video characteristics. We demonstrate strong quantitative results and visually appealing synthesized videos on the BUSV benchmark. Notably, training video classification models on combinations of real and LDDM-synthesized videos substantially improves performance over using real data alone, indicating our method successfully emulates dynamics critical for discrimination. Our image-to-video approach provides an effective data augmentation solution to advance ultrasound video analysis. Code is available at https://github.com/MedAITech/U_I2V.

Ultrasound Image-to-Video Synthesis via Latent Dynamic Diffusion Models

TL;DR

The paper tackles the lack of labeled ultrasound video data by proposing a latent dynamic diffusion model (LDDM) that converts static ultrasound images into plausible dynamic videos. This two-stage approach first encodes video dynamics into a low-dimensional latent space via a video autoencoder and then uses a conditional diffusion model conditioned on the seed image to generate realistic latent trajectories, which are decoded into videos. Evaluations on BUSV and BUSI show that including LDDM-generated videos improves downstream ultrasound video classification and that LDDM outperforms prior image-to-video methods on several quality metrics, while remaining computationally efficient. The method provides a practical data augmentation tool for medical video analysis, and the authors release code to enable adoption and further development.

Abstract

Ultrasound video classification enables automated diagnosis and has emerged as an important research area. However, publicly available ultrasound video datasets remain scarce, hindering progress in developing effective video classification models. We propose addressing this shortage by synthesizing plausible ultrasound videos from readily available, abundant ultrasound images. To this end, we introduce a latent dynamic diffusion model (LDDM) to efficiently translate static images to dynamic sequences with realistic video characteristics. We demonstrate strong quantitative results and visually appealing synthesized videos on the BUSV benchmark. Notably, training video classification models on combinations of real and LDDM-synthesized videos substantially improves performance over using real data alone, indicating our method successfully emulates dynamics critical for discrimination. Our image-to-video approach provides an effective data augmentation solution to advance ultrasound video analysis. Code is available at https://github.com/MedAITech/U_I2V.

Paper Structure

This paper contains 14 sections, 3 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Video Embedding
  4. Conditional Diffusion Model
  5. Video Synthesis from Single Images
  6. Experiments
  7. Datasets
  8. Implement Details
  9. Evaluation Metrics
  10. Results
  11. Video Generation
  12. Synthesized Videos for Classification
  13. Conclusion
  14. Disclosure of Interests.

Figures (3)

  • Figure 1: Overview of the proposed LDDM framework.
  • Figure 2: Visual examples of ultrasound video generation using the proposed LDDM framework. The bottom row shows synthesized video frames generated by our model conditioned on initial frames from the BUSV dataset, which are unseen during training. The top row depicts ground truth frames for reference. As evident from the visual examples, our LDDM approach is capable of generating realistic ultrasound video sequences that are analogous to real ones.
  • Figure 3: Visual examples of generated ultrasound videos based on images from the BUSI dataset.