Lelihu/DR_Artifacts

--- license: mit task_categories: - image-classification - image-segmentation - image-feature-extraction language: - en tags: - medical --- # DDR-Augmented-Artifacts ## Dataset Summary **DDR-Augmented-Artifacts** provides fundus images augmented with realistic synthetic artifacts. Artifacts were cropped from anonymized retina images showing reflections from blood vessels, segmented, and overlaid on DDR images using Gaussian feathered masks and Poisson blending. --- ## Example Below is a sample visualization of how the dataset looks: | Original DDR Image | Augmented with Artifact | |--------------------|--------------------------| | <img src="https://huggingface.co/datasets/shubham212/DR_Artifacts/resolve/main/images/original.jpg" width="250"/> | <img src="https://huggingface.co/datasets/shubham212/DR_Artifacts/resolve/main/images/augmented.png" width="250"/> | --- ## Code Repository The dataset is accompanied by code for artifact generation, preprocessing, and training, available at: 👉 [GitHub Repository](https://github.com/Shubham2376G/DR_Artifacts) This repository contains: - Scripts for generating synthetic artifacts - Example U-Net model for artifact removal ## Supported Tasks - **Image Classification:** Train and evaluate DR classifiers on artifact-rich data. - **Image Segmentation:** Evaluate lesion/DR segmentation robustness under artifacts. - **Preprocessing / Artifact Removal:** Train models to **identify and remove imaging artifacts** prior to downstream analysis. --- ## Languages - Image-based dataset (no natural language component). --- ## Dataset Structure ### Data Fields - `image_id`: filename of the image (string) - `severity_level`: integer label (0–4) indicating DR severity - **0** = No DR - **1** = Mild - **2** = Moderate - **3** = Severe - **4** = Proliferative DR ### Data Splits The dataset is provided as a flat collection of images with a CSV label file. Users should generate **train/val/test splits** at the **patient level** to prevent data leakage. --- ## Intended Uses - Research on **artifact robustness** in medical imaging AI. - Developing augmentation pipelines for retinal image datasets. - Training preprocessing modules to **detect and remove acquisition artifacts**. --- ## Limitations - Synthetic artifacts may not capture full variability of real-world imaging conditions. - Not intended for clinical use. --- ## Ethics and Privacy - All patches were derived from fully anonymized personal retinal images. - No patient-identifiable data is present. - Dataset is for research purposes only. --- ## Citation Please cite both the DDR dataset and this work: ```bibtex @article{LI2019, title = "Diagnostic Assessment of Deep Learning Algorithms for Diabetic Retinopathy Screening", author = "Tao Li and Yingqi Gao and Kai Wang and Song Guo and Hanruo Liu and Hong Kang", journal = "Information Sciences", volume = "501", pages = "511 - 522", year = "2019", issn = "0020-0255", doi = "https://doi.org/10.1016/j.ins.2019.06.011", url = "http://www.sciencedirect.com/science/article/pii/S0020025519305377", } @misc{Aggarwal2025_arxiv, title = DDR-Augmented-Artifacts: Synthetic Artifact Overlays for Robust Diabetic Retinopathy Models, author = Shubham Aggarwal, year = "2025", url = https://arxiv.org/abs/XXXX.XXXXX }许可证：MIT许可证 任务类别： - 图像分类 - 图像分割 - 图像特征提取 语言： - 英语 标签： - 医疗 # DDR-Augmented-Artifacts ## 数据集摘要 **DDR-Augmented-Artifacts** 提供了叠加了逼真合成伪影的眼底图像。此类伪影从已匿名化的视网膜图像中裁剪获取，这些图像包含血管反射痕迹，经分割处理后，通过高斯羽化掩码（Gaussian feathered masks）与泊松融合（Poisson blending）技术叠加至DDR图像之上。 --- ## 示例 以下为该数据集的可视化示例： | 原始DDR图像 | 添加伪影后的图像 | |--------------------|--------------------------| | <img src="https://huggingface.co/datasets/shubham212/DR_Artifacts/resolve/main/images/original.jpg" width="250"/> | <img src="https://huggingface.co/datasets/shubham212/DR_Artifacts/resolve/main/images/augmented.png" width="250"/> | --- ## 代码仓库 本数据集配套了用于伪影生成、预处理与模型训练的代码，开源地址如下： 👉 [GitHub仓库](https://github.com/Shubham2376G/DR_Artifacts) 该仓库包含： - 合成伪影生成脚本 - 用于伪影移除的U-Net模型示例 ## 支持任务 - **图像分类**：在富含伪影的数据上训练并评估糖尿病视网膜病变（Diabetic Retinopathy, DR）分类器。 - **图像分割**：评估模型在伪影干扰下对病变/DR分割的鲁棒性。 - **预处理/伪影移除**：训练模型在开展下游分析前**识别并移除成像伪影**。 --- ## 语言说明 本数据集基于图像构建，无自然语言内容。 --- ## 数据集结构 ### 数据字段 - `image_id`：图像文件名（字符串类型） - `severity_level`：表示DR严重程度的整数标签（取值范围0–4）： - 0 = 无DR - 1 = 轻度病变 - 2 = 中度病变 - 3 = 重度病变 - 4 = 增殖性DR ### 数据划分 本数据集以扁平化图像集合的形式提供，附带CSV格式的标签文件。**用户需按照患者级别生成训练/验证/测试划分**，以避免数据泄露。 --- ## 预期用途 - 医学影像AI领域的**伪影鲁棒性**相关研究 - 为视网膜图像数据集开发数据增强流水线 - 训练预处理模块以**检测并移除采集伪影**。 --- ## 局限性 - 合成伪影无法完全覆盖真实临床成像条件下的全部变异性。 - 本数据集不应用于临床场景。 --- ## 伦理与隐私声明 - 所有伪影补丁均来自完全匿名化的个人视网膜图像。 - 数据集中不包含任何可识别患者身份的信息。 - 本数据集仅用于科研用途。 --- ## 引用要求 请同时引用DDR数据集与本研究： bibtex @article{LI2019, title = "Diagnostic Assessment of Deep Learning Algorithms for Diabetic Retinopathy Screening", author = "Tao Li and Yingqi Gao and Kai Wang and Song Guo and Hanruo Liu and Hong Kang", journal = "Information Sciences", volume = "501", pages = "511 - 522", year = "2019", issn = "0020-0255", doi = "https://doi.org/10.1016/j.ins.2019.06.011", url = "http://www.sciencedirect.com/science/article/pii/S0020025519305377", } @misc{Aggarwal2025_arxiv, title = DDR-Augmented-Artifacts: Synthetic Artifact Overlays for Robust Diabetic Retinopathy Models, author = Shubham Aggarwal, year = "2025", url = https://arxiv.org/abs/XXXX.XXXXX }