Depthwise Separable Convolutional Neural Network Model for Intra-Retinal Cyst Segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Computer Science

Background:

• Intra-retinal cysts (IRCs) are key indicators of various ocular and retinal diseases.
• Accurate segmentation of IRCs from optical coherence tomography (OCT) scans is crucial but challenging due to noise and vendor-specific variations.
• Existing segmentation methods struggle with the complexity and variability of OCT data.

Purpose of the Study:

• To develop a robust convolutional neural network (CNN) model for segmenting intra-retinal cysts (IRCs) across diverse OCT imaging vendors.
• To enhance model generalizability and prevent overfitting in deep learning models for retinal image analysis.
• To improve the accuracy and efficiency of IRC quantification for diagnosing retinal pathologies.

Main Methods:

• Proposed a CNN model featuring an encoder-decoder architecture for IRC segmentation.
• Utilized depthwise separable convolutional filters to reduce computational complexity and improve model generalizability.
• Incorporated the swish activation function to mitigate the vanishing gradient problem.
• Evaluated the model on the Optima Cyst Segmentation Challenge (OCSC) dataset, comprising scans from four different OCT vendors.

Main Results:

• The proposed CNN model achieved a mean Dice score of 0.74 for IRC segmentation.
• The model obtained mean recall and precision rates of 0.72 and 0.82, respectively, across different imaging vendors.
• The developed method demonstrated superior performance compared to existing algorithms on the OCSC dataset.

Conclusions:

• The novel CNN model effectively segments intra-retinal cysts (IRCs) from cross-vendor OCT scans, addressing challenges of noise and intensity variations.
• Depthwise separable convolutions and swish activation contribute to improved model generalizability and performance in retinal image analysis.
• This approach offers a promising tool for accurate IRC quantification, aiding in the diagnosis and management of retinal diseases.