Description

Series information

CORN-1

CORN-1 contains a total of 1698 CCM images of corneal subbasal epithelium using a Heidelberg Retina Tomograph equipped with a Rostock Cornea Module (HRT-III) microscope.  Each image has a resolution of 384 × 384 pixels covering a FOV of 400 × 400 μm2. The manual annotations of the nerve fibres (centerline) in these two datasets were traced by an ophthalmologist using the open source software ImageJ (120  original images were acquired by BioImLab, University of Padova. However,  the manual annotation was made by our ophthalmologist). 

Code

Series information

CORN-2

CORN-2 was divided into low- and high-quality image domains for confocal image enhancement tasks.  A clinical expert was invited and selected 340 low- and 288 high-quality images in this dataset, respectively, for training and 60 low-quality images for testing. For evaluation by nerve fiber segmentation, we also provided manual annotation of nerve fibers at centerline pixel level in the 60 testing images. All the images of CORN-2 were acquired at a size 384 × 384.

Code

Series information

CORN-3

For automated tortuosity analysis of nerve fibers in corneal confocal microscopy, we constructed CORN-3 based on CCM-A. CORN-3 was categorized into four groups based on fiber tortuosity level. An image analysis expert and the clinical author (obs 1 and obs 2) each independently labeled the tortuosity level according to a previously published protocol, and the consensus between them was then used as ground truth (GT), i.e., Level 1: the fibers appear almost straight (54 images); Level 2: the fibers appear moderately tortuous (212 images); Level 3: the fibers are quite tortuous; the amplitude of the changes in the fiber direction is quite severe (108 images); Level 4: the fibers appear very tortuous, presenting frequent changes in the fiber direction (29 images). All the images of CORN-3 were acquired at a size 384 × 384.

Series information

CORN¹⁵⁰⁰

CORN¹⁵⁰⁰ is a new corneal nerve dataset which was particularly designed for tortuosity grading tasks. The newly constructed CORN¹⁵⁰⁰ aims to expand the dataset pool for tortuosity grading. Similar to CORN-3, all 1500 CCM in CORN¹⁵⁰⁰ were acquired using a Heidelberg Retina Tomograph equipped with a Rostock Cornea Module (HRT-III) microscope. Each image has a resolution of 384 × 384 pixels covering a field of view of 400 × 400um². All images were manually annotated by the same ophthalmologist as before into four tortuosity levels (level 1 to level 4) and divided into a training set (1200 images) and testing set (300 images).

Series information

CORN-Pro

To the best of our knowledge, no publicly available dataset exists for the segmentation of Langerhans cells (LCs) or stromal cells (SCs). To address this gap, we have extended the CORN dataset by introducing a new subset, CORN-4, specifically designed for segmentation tasks involving LCs and SCs. CORN-4 consists of 1,120 corneal confocal microscopy (CCM) images captured using the Heidelberg Retina Tomography System with the Rostock Cornea Module (HRT-III). The dataset includes 560 images featuring both corneal nerves and LCs, and another 560 images displaying corneal nerves and/or stromal cells. All images were acquired at a distance of 40-60 μm from the anterior surface of the central cornea, with a resolution of 384 × 384 pixels, covering an area of 400 × 400 μm. Manual annotation was carried out by an image analysis expert using the open-source software ITK-SNAP, and the labels were subsequently reviewed and corrected by a senior ophthalmologist.

Series information

CORN-Complex

CORN-Complex is a large-scale, multi-center, multi-disease dataset designed to reflect real-world clinical complexity. It comprises 7,522 images from 518 patients across 12 distinct corneal pathologies. Data were collected from six clinical centers using two imaging systems: Confoscan 3.0 (resized to 384×384) and Heidelberg HRT-III/RCM (384×384). All images were exported directly from CCM devices without manual quality filtering to preserve the full spectrum of variability encountered in clinical practice. A representative subset of 1,429 images, ensuring balanced coverage of all disease categories, imaging devices, and clinical centers, was exhaustively annotated by two image analysis experts and reviewed by a senior ophthalmologist, providing high-quality ground truth labels for corneal nerves (CNs) and Langerhans cells (LCs). To ensure reliable benchmarking, we define a fixed data split: 5% of the total dataset (376 images) is allocated as the test set, with another 5% (376 images) reserved for validation. The remaining 6,770 images form the training pool, of which 10% (677 images) are labeled.

12 corneal pathologies: Fungal keratitis; Viral keratitis; Anterior Basement Membrane Dystrophy; Neurotrophic Keratitis; Nonspecific keratitis; Keratoconus; Fuchs endothelial dystrophy; Dry eye disease; Post-corneal surgery; Neuropathic corneal pain; Diabetes mellitus; Corneal endotheliitis.