Scholarlink

<p>Early detection of vitreoretinal surface disorders is crucial, as they constitute significant factors contributing to the progression of macular degeneration. Despite their clinical importance, vitreomacular adhesion and traction (VMA-VMT) stages are almost absent in existing public OCT (optical coherence tomography) datasets, representing a significant gap in the development of artificial intelligence (AI) based diagnostic tools. Most publicly available OCT datasets include images of standard classes (diseases) like age-related macular degeneration (AMD), choroidal neovascularization (CNV), diabetic macular edema (DME), drusen (DRUSEN), and normal retina (NORMAL). In this article, we introduce a vitreo-macular adhesion or traction (VMA-VMT) class of B-scans acquired using the Heidelberg Spectralis device.To maintain the consistency of image parameters (including histograms), the dataset was also prepared using subsets of other classes. The CAVRI-H5 dataset comprises a total of 609 high-quality OCT B-scans. The images are distributed across five classes: NORMAL (142), DRUSEN (137), VMA-VMT (135), DME (99), and CNV (96). For model evaluation, we employed a stratified split, allocating 10 images from each class to the validation and test sets, with the remaining images used for training. All images, categorized by qualified ophthalmologists, provide a unique resource for deep learning classification. The resulting expert-annotated dataset provides a valuable resource for deep learning-based retinal OCT classification. Technical validation using six neural network architectures, including traditional CNNs (Convolutional Neural Networks), efficient multiscale CNNs, and a Vision Transformer-based model (ViT), demonstrated the suitability of CAVRI-H5 for deep learning-based retinal OCT classification.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CAVRI-H5: A Homogeneous OCT Dataset for Retinal Classification Featuring VMA-VMT Pathologies

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月5日

查看全文
  • Alicja Ignatowicz,
  • Tomasz Marciniak,
  • Agnieszka Stankiewicz,
  • Marcin Stopa,
  • Elżbieta Marciniak

摘要

Early detection of vitreoretinal surface disorders is crucial, as they constitute significant factors contributing to the progression of macular degeneration. Despite their clinical importance, vitreomacular adhesion and traction (VMA-VMT) stages are almost absent in existing public OCT (optical coherence tomography) datasets, representing a significant gap in the development of artificial intelligence (AI) based diagnostic tools. Most publicly available OCT datasets include images of standard classes (diseases) like age-related macular degeneration (AMD), choroidal neovascularization (CNV), diabetic macular edema (DME), drusen (DRUSEN), and normal retina (NORMAL). In this article, we introduce a vitreo-macular adhesion or traction (VMA-VMT) class of B-scans acquired using the Heidelberg Spectralis device.To maintain the consistency of image parameters (including histograms), the dataset was also prepared using subsets of other classes. The CAVRI-H5 dataset comprises a total of 609 high-quality OCT B-scans. The images are distributed across five classes: NORMAL (142), DRUSEN (137), VMA-VMT (135), DME (99), and CNV (96). For model evaluation, we employed a stratified split, allocating 10 images from each class to the validation and test sets, with the remaining images used for training. All images, categorized by qualified ophthalmologists, provide a unique resource for deep learning classification. The resulting expert-annotated dataset provides a valuable resource for deep learning-based retinal OCT classification. Technical validation using six neural network architectures, including traditional CNNs (Convolutional Neural Networks), efficient multiscale CNNs, and a Vision Transformer-based model (ViT), demonstrated the suitability of CAVRI-H5 for deep learning-based retinal OCT classification.