ARTIFICIAL INTELLIGENCE'S ROLE IN DIFFERENTIATING THE ORIGIN FOR SUBRETINAL BLEEDING IN PATHOLOGIC MYOPIA.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Pathologic myopia can cause subretinal hemorrhage (SH), which may stem from choroidal neovascularization (CNV) or simple bleeding (SB).
• Accurate differentiation is crucial for appropriate treatment and management.
• Distinguishing between these causes using optical coherence tomography (OCT) B-scan images can be challenging.

Purpose of the Study:

• To identify key imaging features for differentiating SH due to CNV onset from SH without CNV (SB) in pathologic myopia.
• To evaluate the utility of a machine learning (ML)-based approach in supporting differential diagnosis.

Main Methods:

• Employed four feature extraction methods: GradCAM visualization, reverse engineering, image processing, and human grader measurements.
• Utilized a deep learning model (Inception-ResNet-v2) for GradCAM analysis on OCT B-scan images.
• Compared diagnostic performance of human graders with and without ML-derived feature guidance.

Main Results:

• CNV eyes exhibited higher baseline central macular thickness.
• Image processing revealed subretinal material inhomogeneity and Bruch membrane interruption at SH margins in CNV eyes.
• Human grader accuracy improved from 76.9% to 83.3% with ML guidance (P=0.02).
• Deep learning model achieved 86.0% accuracy in classifying SH causes.

Conclusions:

• Artificial intelligence can identify imaging biomarkers suggestive of CNV in SH contexts within myopic eyes.
• AI enhances human diagnostic capabilities for differentiating SH causes using baseline OCT images.
• Deep learning models demonstrate high accuracy in distinguishing between CNV-related and simple bleeding SH.