Efficient Semantic Diffusion Architectures for Model Training on Synthetic Echocardiograms Dataset

This is the official data repository for the paper: "Efficient Semantic Diffusion Architectures for Model Training on Synthetic Echocardiograms", available at: https://www.arxiv.org/abs/2409.19371. The corresponding code is available at: https://github.com/david-stojanovski/echo_from_noise

The synthetic data is produced using a variety of generative architectures, including the Elucidating Diffusion Model (EDM), Variance Exploding (VE), Variance Preserving (VP), and our novel models, EDM-L64 and EDM-L128, which employ latent diffusion strategies to significantly reduce computational cost. By incorporating spatially adaptive normalization (SPADE) blocks and Γ-distribution-based Variational Autoencoders (Γ-VAE), these datasets ensure that the generated images preserve the essential semantic features required for training deep learning models.

All pretrained classification and segmentation models can be found within the trained_models file.

All generated images can be found within the generated_data file. Included is the CAMUS and original Semantic Diffusion Model (SDM) data, as well as a folder labelled easy_inference designed to contain all relevant labelmaps in a convenient folder for generating replicas of the dataset (detailed at codebase).

本仓库为论文《用于合成超声心动图模型训练的高效语义扩散架构》（Efficient Semantic Diffusion Architectures for Model Training on Synthetic Echocardiograms）的官方数据集仓库，论文公开链接：https://www.arxiv.org/abs/2409.19371。配套代码开源地址：https://github.com/david-stojanovski/echo_from_noise

本合成数据集依托多种生成架构构建，涵盖阐明扩散模型（Elucidating Diffusion Model, EDM）、方差爆炸（Variance Exploding, VE）、方差保持（Variance Preserving, VP），以及我们提出的新型模型EDM-L64与EDM-L128。上述模型采用潜在扩散策略，可显著降低训练计算开销。通过融入空间自适应归一化（spatially adaptive normalization, SPADE）模块与基于Γ分布的变分自编码器（Γ-VAE），本数据集生成的图像可保留深度学习模型训练所需的核心语义特征。

所有预训练分类与分割模型均存放于trained_models文件夹中。

所有生成图像均存放于generated_data文件夹内，其中包含CAMUS数据集与原始语义扩散模型（Semantic Diffusion Model, SDM）数据，同时设有名为easy_inference的文件夹，该文件夹以便捷的目录结构存储所有相关标签图，用于复现本数据集（具体细节见配套代码库）。