Scholarlink

<p>To develop a deep learning–based computer-aided diagnostic model for the automated identification of corneal microneuromas from in vivo confocal microscopy (IVCM) images and to preliminarily assess its potential clinical utility in the context of diabetic corneal neuropathy. This retrospective diagnostic accuracy study was conducted at the Ophthalmology Center of Renmin Hospital of Wuhan University from August 2021 to December 2023. A total of 293 patients (358 eyes) underwent IVCM, yielding 4,554 corneal images, which were divided into a training set (3,542 images) and an internal test set (1,012 images). For external validation, an independent cohort of 87 patients (103 eyes) contributed 506 IVCM images from Wuhan Aier Hankou Eye Hospital. All images underwent standardized quality control and graded annotation by three senior corneal specialists prior to model development. The modeling pipeline included automated image-quality assessment and non–target-layer filtering, microneuroma detection, and microneuroma subtype classification. Model generalizability and potential clinical applicability were evaluated using the independent external dataset. The diagnostic model achieved an overall accuracy of 0.975 in image-quality screening. For identifying the presence or absence of microneuromas, the overall accuracy was 0.837 on the internal test set and 0.812 on the external test set. The accuracy range for the three morphological subtypes was 0.798–0.886. In the image-reading experiment, assistance from the model significantly improved the junior physicians’ accuracy from 0.687 to 0.875 (<i>P</i> &lt; 0.001) and increased their reading efficiency 2.2-fold, reducing the average reading time from 63.4 ± 15.2&#xa0;s to 28.5 ± 6.8&#xa0;s (<i>t</i> = 19.3, <i>P</i> &lt; 0.001). These findings reflect model performance within the available datasets. We developed a deep learning–based system capable of automatically analyzing IVCM images to detect corneal microneuromas. The model demonstrated strong and consistent performance across internal and external datasets and substantially enhanced the diagnostic accuracy and efficiency of junior clinicians.</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

AI-Assisted detection of corneal nerve structural abnormalities in early diabetic keratopathy: development and validation of a deep learning framework

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月20日

查看全文
  • Jiali Pan,
  • Xiaoshuo Shi,
  • Longbiao Wan,
  • Lan Ke,
  • Yulin Yan,
  • Jiewen Mao,
  • Yuyu Cong,
  • Zixian Yang,
  • Juntao Fang,
  • Yingli Wang,
  • Chenhao Xu,
  • Minghui Bao,
  • Biqing Zheng,
  • Ling Gao,
  • Linying Huang,
  • Yanning Yang

摘要

To develop a deep learning–based computer-aided diagnostic model for the automated identification of corneal microneuromas from in vivo confocal microscopy (IVCM) images and to preliminarily assess its potential clinical utility in the context of diabetic corneal neuropathy. This retrospective diagnostic accuracy study was conducted at the Ophthalmology Center of Renmin Hospital of Wuhan University from August 2021 to December 2023. A total of 293 patients (358 eyes) underwent IVCM, yielding 4,554 corneal images, which were divided into a training set (3,542 images) and an internal test set (1,012 images). For external validation, an independent cohort of 87 patients (103 eyes) contributed 506 IVCM images from Wuhan Aier Hankou Eye Hospital. All images underwent standardized quality control and graded annotation by three senior corneal specialists prior to model development. The modeling pipeline included automated image-quality assessment and non–target-layer filtering, microneuroma detection, and microneuroma subtype classification. Model generalizability and potential clinical applicability were evaluated using the independent external dataset. The diagnostic model achieved an overall accuracy of 0.975 in image-quality screening. For identifying the presence or absence of microneuromas, the overall accuracy was 0.837 on the internal test set and 0.812 on the external test set. The accuracy range for the three morphological subtypes was 0.798–0.886. In the image-reading experiment, assistance from the model significantly improved the junior physicians’ accuracy from 0.687 to 0.875 (P < 0.001) and increased their reading efficiency 2.2-fold, reducing the average reading time from 63.4 ± 15.2 s to 28.5 ± 6.8 s (t = 19.3, P < 0.001). These findings reflect model performance within the available datasets. We developed a deep learning–based system capable of automatically analyzing IVCM images to detect corneal microneuromas. The model demonstrated strong and consistent performance across internal and external datasets and substantially enhanced the diagnostic accuracy and efficiency of junior clinicians.