Optimising deep learning models for ophthalmological disorder classification.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Fundus imaging is crucial for detecting ophthalmological diseases like glaucoma and diabetic retinopathy.
• Structural changes in the optic disc, macula, and blood vessels signal various eye conditions.
• Accurate diagnosis of these conditions is vital for patient eye health.

Purpose of the Study:

• To apply deep learning models for multi-class and multi-label classification of ophthalmological disorders.
• To evaluate the effectiveness of transfer learning-based convolutional neural network (CNN) methods.
• To compare the performance of different optimizers for improved diagnostic accuracy.

Main Methods:

• Utilized transfer learning with CNN models for image classification.
• Employed the Ocular Disease Intelligent Recognition (ODIR) database containing fundus images.
• Compared Stochastic Gradient Descent (SGD) and Adam optimizers.

Main Results:

• The MobileNet model with the Adam optimizer achieved the highest testing accuracy.
• Achieved a testing accuracy of 89.64% for ophthalmological disorder classification.
• Demonstrated the efficacy of deep learning in diagnosing multiple eye conditions.

Conclusions:

• Deep learning, particularly CNNs, shows significant promise in classifying ophthalmological disorders.
• The MobileNet architecture combined with the Adam optimizer offers a robust approach for fundus image analysis.
• This methodology can aid in the early detection and management of eye diseases.