Analysis of anterior segment in primary angle closure suspect with deep learning models

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Primary angle closure suspect (PACS) is a condition characterized by anatomical variations in the anterior chamber configuration.
• Early detection and screening of PACS are crucial for preventing vision loss.
• Traditional diagnostic methods rely on manual examination and may be limited in efficiency and accuracy.

Purpose of the Study:

• To analyze the distinct anatomical characteristics of the anterior chamber in PACS patients.
• To develop and evaluate an artificial intelligence (AI)-aided diagnostic system for effective PACS screening.

Main Methods:

• A cross-sectional study involving 1668 anterior segment optical coherence tomography (AS-OCT) scans from 839 patients (PACS group and normal group).
• Comparison of anatomical features between PACS and normal groups.
• Construction and evaluation of AI models (CART, RF, LR, VGG-16, Alexnet) for PACS screening, comparing their diagnostic efficiency against junior physicians and experienced ophthalmologists.

Main Results:

• Random Forest (RF) and Classification and Regression Tree (CART) algorithms demonstrated high performance in PACS screening (AUC=0.90).
• Convolutional Neural Networks (CNNs), particularly Alexnet, also showed strong diagnostic capabilities (Alexnet: Se=0.83, Sp=0.95, PPV=0.92, NPV=0.87, AUC=0.85).
• The AI models, especially CNNs, outperformed junior physicians and achieved diagnostic performance comparable to experienced ophthalmologists.

Conclusions:

• PACS patients possess identifiable anatomical characteristics distinct from healthy individuals.
• AI-driven screening models are a reliable and potent tool for PACS detection.
• The developed AI system demonstrates diagnostic equivalence to that of experienced ophthalmologists, offering a promising advancement in ophthalmic screening.