Automatic Segmentation and Measurement of Choroid Layer in High Myopia for OCT Imaging Using Deep Learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate choroid layer segmentation is crucial for diagnosing fundus diseases like diabetic retinopathy and high myopia.
• Existing segmentation algorithms struggle with the blurred boundaries and complex gradients of choroidal layers.
• Optical coherence tomography (OCT) is a key imaging modality for visualizing the choroid.

Purpose of the Study:

• To develop a novel choroid segmentation method combining image enhancement and an attention-based dense U-Net (AD-U-Net).
• To evaluate the performance of the proposed method in segmenting choroid layers from OCT images of normal and high myopia eyes.
• To introduce algorithms for automatic measurement of choroidal foveal thickness and adjacent volumes.

Main Methods:

• Pre-enhancement of OCT images using flattening, filtering, and exponential/linear enhancement.
• Implementation of an attention-based dense U-Net (AD-U-Net) for choroid layer segmentation.
• Validation using 800 OCT B-scans from normal and high myopia subjects.

Main Results:

• Image enhancement significantly improved AD-U-Net performance, achieving an Area Under the Curve (AUC) of 99.51% and a Dice Similarity Coefficient (DSC) of 97.91%.
• The AD-U-Net method demonstrated superior segmentation accuracy compared to existing networks.
• High myopia eyes showed an 86.3% reduction in choroidal foveal thickness and a 90% reduction in adjacent volume compared to normal eyes.

Conclusions:

• The proposed image enhancement and AD-U-Net method offers a highly accurate approach for choroid layer segmentation in OCT images.
• High myopia is strongly associated with significant choroid layer attenuation.
• These algorithms hold potential for the diagnosis and prevention of fundus lesions related to choroid thinning in high myopia.