Deep learning-based estimation of axial length using macular optical coherence tomography images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Axial length (AL) estimation is critical for diagnosing and managing various eye conditions.
• Macular optical coherence tomography (OCT) images offer detailed retinal visualization.
• Developing non-invasive methods for AL estimation is an ongoing research area.

Purpose of the Study:

• To develop and evaluate deep learning models for estimating axial length (AL) using macular OCT images.
• To compare the performance of deep learning models against traditional methods based on retinal thickness.
• To identify specific macular features associated with axial length.

Main Methods:

• Utilized 2,664 macular OCT images from 444 patients without maculopathy.
• Developed and trained three pre-trained deep learning models (ResNet 18, ResNet 50, ViT) for binary classification and regression tasks.
• Employed ten-fold cross-validation and Grad-CAM analysis for model evaluation and feature visualization.

Main Results:

• ResNet 50 achieved high accuracy (87.2%) and AUC (0.929) for AL estimation from OCT images.
• Deep learning models significantly outperformed retinal thickness measurements (AUC 0.747, accuracy 77.8%).
• Grad-CAM analysis localized AL-related macular features to the fovea and adjacent macular regions.

Conclusions:

• Macular OCT images, analyzed with deep learning, provide an effective method for estimating axial length.
• The identified localized macular features offer insights into the pathogenesis of AL-related maculopathy.
• This approach lays the groundwork for future research into AL-related retinal diseases.