TL-MED: Multiclass eye disease classification based on ensemble transfer learning and CRVO-BRVO detection via a...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Imaging

Background:

• Traditional methods for diagnosing eye diseases have limitations.
• Deep convolutional neural networks (DCNNs) show promise in medical image analysis.
• Transfer learning can enhance the performance of DCNNs.

Purpose of the Study:

• To develop and evaluate state-of-the-art DCNN models for diagnosing common eye diseases.
• To investigate the effectiveness of transfer learning in improving diagnostic accuracy.
• To create an ensemble model for superior performance in eye disease classification.

Main Methods:

• Trained five individual DCNN models (VGG16, ResNet152, DenseNet169, EfficientNetB3, NASNetMobile) using transfer learning on 3744 retinal images.
• Developed an ensemble model combining ResNet152, DenseNet169, and EfficientNetB3.
• Utilized a single-shot multibox detector (SSD) for detecting retinal vein occlusions.

Main Results:

• DenseNet169 achieved 96% accuracy, while NASNetMobile achieved 87% accuracy.
• The ensemble model outperformed individual models, reaching a peak accuracy of 97%.
• Results demonstrate the effectiveness of transfer learning for robust eye disease classification.

Conclusions:

• AI-driven approaches using DCNNs offer a reliable solution for early and precise detection of eye diseases.
• The proposed models can assist healthcare professionals in diagnosis, leading to improved patient care.
• This work contributes to the development of advanced diagnostic tools in ophthalmology.