# Validated semi-supervised early and accurate screening for anterior segment diseases: a 3PM-guided conceptual and technological innovation

## TL;DR

This study introduces a validated semi-supervised AI system for early and accurate detection of anterior segment eye diseases, improving screening efficiency and accuracy.

## Contribution

A novel semi-supervised object detection framework with modules to handle class imbalance and detect unseen lesions in slit-lamp imaging.

## Key findings

- The SSOD achieved comparable mAP to YOLOv8 but significantly higher recall for both single- and multi-lesion detection.
- In clinical evaluations, SSOD outperformed YOLOv8 and approached junior ophthalmologists in multi-lesion detection.
- The system supports early lesion recognition and individualized ophthalmic care aligned with 3PM principles.

## Abstract

Ocular anterior segment diseases are major causes of global visual impairment. Early and accurate detection of anterior segment abnormalities is essential to support predictive diagnostics, targeted prevention, and individualized treatments management. Conventional slit-lamp assessments are often limited by human observation and inter-clinician variability, restricting their ability to achieve rapid and large-scale disease screening. To advance anterior segment care within the predictive, preventive, and personalized medicine (PPPM/3PM) framework, this study aimed to develop a comprehensive and validated semi-supervised object detection (SSOD) system for slit-lamp imaging–based screening of multiple anterior segment diseases.

A total of 7230 slit-lamp images from 3302 patients were retrospectively collected at the Second Affiliated Hospital of Zhejiang University between November 2016 and July 2024. The proposed SSOD integrated a Category Control Embed (CCE) module to mitigate class imbalance and an Out-of-distribution Detection Fusion Classifier (ODDFC) to identify previously unseen lesions. Model performance was quantitatively compared with YOLOv8 and ophthalmologists using quantitative metrics (average precision [AP], recall) and clinical assessments of diagnostic accuracy, lesion comprehensiveness, and localization precision.

The SSOD achieved mAP comparable to YOLOv8 (0.729 vs. 0.725 for single-lesion; 0.538 vs. 0.543 for multi-lesion), but demonstrated substantially higher recall (0.893 vs. 0.656 for single-lesion; 0.679 vs. 0.477 for multi-lesion). In clinical evaluations, SSOD scored 2.430/3 for single-lesion and 1.942/3 for multi-lesion detection, outperforming YOLOv8 and approaching the performance of junior ophthalmologists in multi-lesion cases.

The SSOD framework offers an efficient and scalable solution for anterior segment disease screening, delivering reliable multi-lesion detection with minimal annotation. It supports early recognition of anterior segment lesions, guides targeted interventions to prevent irreversible vision loss, and facilitates patient-centered, individualized management that advances ophthalmic care from reactive assessment to proactive precision treatment, aligning with the principles of 3PM.