InfoMotion: A Graph-Based Approach to Video Dataset Distillation for Echocardiography

Zhe Li, Hadrien Reynaud, Alberto Gómez, Bernhard KainzarXiv 2025 · DOI · arXiv

Introduction

Deep learning has significantly advanced echocardiography, particularly in assessing the Ejection Fraction (EF)—the primary metric for evaluating cardiac pump function. However, the field faces two critical challenges: the massive volume of video data, which creates storage and computational bottlenecks, and strict privacy regulations (e.g., HIPAA, GDPR) that hinder data sharing.

Dataset Distillation offers a potential solution by synthesizing a small, representative dataset from a large original pool. While effective for 2D images, distilling video is complex due to high frame-to-frame similarity and the need to capture temporal dynamics. InfoMotion is the first framework designed to address these challenges specifically for medical video datasets.

The InfoMotion Framework

The InfoMotion pipeline shifts the focus from static appearance to cardiac motion, ensuring that the distilled samples capture clinically relevant functional patterns.

The InfoMotion pipeline: from motion extraction to graph-based video selection.

Motion Feature Extraction

Rather than relying on standard CNNs that conflate appearance and motion, InfoMotion utilizes Inter-Frame Attention (IFA). This mechanism explicitly isolates motion features, which are more discriminative for heart conditions like heart failure than heart shape alone.

Graph-Based Selection and Infomap

To select the most informative videos, the framework constructs a graph where each video is a node and edges represent motion similarity. The Infomap algorithm—typically used for community detection in complex networks—identifies clusters of similar videos. By selecting videos with high "modular centrality" from each community, the method ensures the final distilled set is both diverse and highly representative of the original data distribution.

Temporal structure of input frames and the resulting graph-based community representation.

Experimental Setup

The study utilized two primary benchmarks:

  • EchoNet-Dynamic: A public dataset of over 10,000 real-world echocardiograms with expert EF labels.
  • EchoNet-Synthetic: A pool of 7,500 videos generated by a latent video diffusion model, from which the distilled set is selected.
  • Latte: A secondary diffusion model used to verify the framework's generalizability across different generative sources.

Key Results and Insights

InfoMotion demonstrates that motion is the primary driver of clinical relevance in ultrasound distillation. Key findings include:

  • Extreme Efficiency: A model trained on just 25 synthetic videos (5 per clinical class) achieved 69.38% accuracy on real patient data. This is a remarkable result compared to the 81.21% accuracy achieved by training on the full dataset of 7,465 videos.
  • Superior Stability: Compared to random selection or K-means clustering, InfoMotion showed significantly lower variance (±2.2% vs. ±17.8% for random), indicating a highly reliable distillation process.
  • Topological Integrity: Using graph-based community detection proved superior to standard clustering because it respects the underlying topology of the medical data, leading to more accurate representative sampling.

Conclusion and Future Work

InfoMotion represents a major step forward in medical video distillation, successfully preserving clinical utility while drastically reducing data requirements. Future research will explore extending this framework to other tasks, such as video segmentation and view classification, and further optimizing the computational efficiency of large-scale graph searches. As generative video models continue to improve, the quality and utility of distilled medical datasets are expected to advance in tandem.

Summary updated

Abstract

Echocardiography plays a critical role in the diagnosis and monitoring of cardiovascular diseases as a non-invasive real-time assessment of cardiac structure and function. However, the growing scale of echocardiographic video data presents significant challenges in terms of storage, computation, and model training efficiency. Dataset distillation offers a promising solution by synthesizing a compact, informative subset of data that retains the key clinical features of the original dataset. In this work, we propose a novel approach for distilling a compact synthetic echocardiographic video dataset. Our method leverages motion feature extraction to capture temporal dynamics, followed by class-wise graph construction and representative sample selection using the Infomap algorithm. This enables us to select a diverse and informative subset of synthetic videos that preserves the essential characteristics of the original dataset. We evaluate our approach on the EchoNet-Dynamic datasets and achieve a test accuracy of \(69.38\%\) using only \(25\) synthetic videos. These results demonstrate the effectiveness and scalability of our method for medical video dataset distillation.
BibTeX
@misc{li2025infomotion,
  author = {Zhe Li and Hadrien Reynaud and Alberto Gómez and Bernhard Kainz},
  title = {{InfoMotion: A Graph-Based Approach to Video Dataset Distillation for Echocardiography}},
  year = {2025},
  eprint = {2512.09422},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.09422},
}
Figure