Deep Ensemble for Central Serous Microscopic Retinopathy Detection in Retinal Optical Coherence Tomographic Images

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Central serous retinopathy (CSR) is a common retinal disorder causing vision loss.
• Manual detection methods for CSR are often imprecise and unreliable.
• Early and accurate detection of CSR is crucial for preventing vision impairment.

Purpose of the Study:

• To develop and evaluate a deep learning-based framework for automated classification of central serous retinopathy (CSR).
• To address challenges in segmenting retinal images, such as noise and variations in fluid characteristics.
• To improve the accuracy and reliability of CSR detection compared to traditional methods.

Main Methods:

• A convolutional neural network (CNN)-based framework was developed, incorporating image segmentation and post-processing techniques.
• Otsu's thresholding was used for segmenting optical coherence tomography (OCT) images, followed by contrast adjustment and noise removal.
• Image classification was performed using a fusion of ResNet-18, GoogleNet, and VGG-19 networks on the OCTID dataset.

Main Results:

• The proposed framework achieved high performance metrics: 99.6% accuracy, 99.46% sensitivity, 100% specificity, and 99.73% F1 score.
• The method successfully classified normal and CSR-affected images from the publicly available OCTID dataset.
• The fusion of three CNNs demonstrated superior performance in identifying CSR.

Conclusions:

• The developed deep learning framework is highly effective and suitable for automated CSR classification.
• The study validates the framework's ability to overcome segmentation challenges in retinal OCT images.
• The findings support the potential of this AI-driven approach for early CSR diagnosis and vision preservation.