SLID: a slit-lamp image dataset for deep learning-based anterior eye anatomical segmentation and multi-lesion detection

Ocular anterior segment diseases are a significant global public health concern, being major contributors to blindness and visual impairment worldwide. [1][2][3][4] Cataract, known as the leading cause of reversible blindness and visual impairment globally, [5][6][7] along with corneal diseases, which rank as the fourth leading cause of blindness, 8,9 underscore the urgent need for timely diagnosis and intervention, particularly for conditions like keratitis that can progress rapidly. 10,11 Among the essential tools for diagnosing and managing these conditions is the slit-lamp biomicroscope, a fundamental instrument in ophthalmology clinics known for its convenience, availability, cost-effectiveness, and efficiency in identifying common ocular anterior segment diseases.Recent advancements in artificial intelligence (AI) have shown promising results in automated diagnosis and treatment planning based on slit-lamp images. [12][13][14][15][16][17][18][19][20][21][22][23] However, these AI systems are often limited by the availability of comprehensive datasets for training and validation purposes. Currently, there is a scarcity of open-access datasets that include slit-lamp images with detailed anatomical annotations and lesion identification necessary for developing robust AI models applicable to realworld clinical scenarios.The publication of slit-lamp image datasets with thorough anatomical annotations and lesion identification is therefore pivotal. These datasets not only facilitate the advancement of AI models capable of precise and clinically relevant diagnoses but also serve to validate existing AI technologies. By offering standardized datasets to researchers globally, we aim to expedite progress toward more efficient computer-aided diagnosis and treatment systems for ocular anterior segment diseases.This paper introduces a comprehensive dataset crafted to fill these critical gaps in the field. Figure 1 illustrates our study, delineating the processes of data collection, anatomical segmentation, and lesion identification. To the best of our knowledge, this is the first open-source slit-lamp dataset with detailed anatomical and lesion-level information. Our dataset endeavors to serve as an invaluable resource for researchers endeavoring to develop AI models that are not only accurate but also applicable in preclinical screening and clinical practice.Data Report 3 Materials and methodsWe retrospectively collected 2617 ocular surface slit-lamp images from 1119 patients between November 2016 and March 2022 at the Eye Center of the Second Affiliated Hospital of Zhejiang University, College of Medicine, China. All images were captured using Topcon SL-D701 slit-lamp biomicroscopes equipped with DC-4 digital cameras. Images were excluded based on the following criteria: (1) Poor quality images, defined as images with significant defocus or inappropriate illumination (overexposed or underexposed conditions) such that lesion features cannot be clearly identified; (2) Indeterminate lesions, defined as cases with atypical or ambiguous presentations in which our annotation team could not reach a consensus on the specific lesion category; (3) Anatomical abnormalities, defined as structural alterations of the ocular surface caused by prior surgery or trauma.To ensure anonymity, patient information was obscured with a black rectangular box in the top left corner of each image. Due to multiple follow-up visits for the same patient and varying eye positions and focus points during the same visit, multiple images may exist for the same eye of a single patient.Demographically, a total of 1383 eyes were included, comprising 668 left eyes and 715 right eyes, corresponding to 1119 patients. Of these, 621 were women and 498 were men. Age distribution was as follows: 236 patients 0.918 and mean Dice > 0.955 across all three regions and all four experts. Notably, segmentation performance for the conjunctival region was slightly superior to that of the corneal and pupillary regions, likely due to the clearer boundary between the conjunctiva and surrounding eyelid skin tissue. The validation results of lesion identification showed acceptable consistency, with mean IoU > 0.764 and mean Dice > 0.835 across all four experts. Among these images, those with monomorbidity exhibited relatively higher consistency, achieving a mean IoU of 0.785 and mean Dice score of 0.861. Images with multimorbidity also showed acceptable consistency, with a mean IoU of 0.717 and mean Dice score of 0.786. When analyzing specific lesion types, pterygium and pinguecula demonstrated lower consistency, primarily due to their less distinct lesion boundaries.To evaluate the potential of the proposed dataset for deep learning-based anterior eye multi-lesion detection, a YOLOv8 model was trained using single-lesion images following standard procedures, with experiment parameters listed in Table S1. 25 At the image level, the dataset was randomly divided into training, validation, and test sets in an 8:1:1 ratio, achieving an average mean Average Precision (mAP) of 0.873 across all 13 single-lesion categories (Table S2). At the patient level, we re-executed the 8:1:1 split, ensuring that all images from the same patient were assigned to a single dataset (training, validation, or test). Under this patient-level split, YOLOv8 achieved an average mAP of 0.736. Data availability De-identified data were used in the present study. These data are publicly accessible at GitHub as of the date of publication.During the review stage, reviewers can download it using a token at GitHub. Specifically:(1) Annotations.csv can be downloaded using the code: curl -H "Authorization: token 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" \ -L "https://api.github.com/repos/xumingyu-hub/SLID/contents/Annotations.csv" \ -o Annotations.csv (2) Original_Slit-lamp_Images.zip can be downloaded using the code: curl -H "Authorization: token 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" \ -L "https://api.github.com/repos/xumingyu-hub/SLID/contents/Original_Slit-lamp_Images.zip"\ -o Original_Slit-lamp_Images.zip (3) After downloading the files, open them in a text editor (e.g., .txt), then copy the URL following "download_url" to successfully download the data.