A Comprehensive Performance Analysis of Transfer Learning Optimization in Visual Field Defect Classification.

Area of Science:

• Ophthalmology
• Computer Science
• Artificial Intelligence

Background:

• Convolutional Neural Network (CNN) models demonstrate high accuracy in classifying visual field (VF) defects.
• Several pre-trained models, including VGG-Net, MobileNet, ResNet, and DenseNet, are suitable for VF defect classification.

Purpose of the Study:

• To compare the performance of different pre-trained CNN models for VF defect classification.
• To evaluate the impact of hyperparameter tuning and fine-tuning on CNN model performance.
• To identify the optimal CNN model and configuration for accurate VF defect classification.

Main Methods:

• Comparative analysis of VGG-Net, MobileNet, ResNet, and DenseNet models.
• Implementation of hyperparameter tuning and fine-tuning using Bayesian optimization.
• Evaluation of model performance using accuracy metrics on validation and test datasets.

Main Results:

• VGG-16 achieved 97.63% accuracy before hyperparameter optimization.
• Automated hyperparameter tuning and fine-tuning significantly enhanced model performance.
• DenseNet-121 demonstrated the best performance with 98.46% validation accuracy and 99.57% test accuracy.

Conclusions:

• Hyperparameter tuning and fine-tuning are crucial for optimizing deep learning models in VF defect classification.
• Bayesian optimization effectively automates the selection of optimal hyperparameters and fine-tuning strategies.
• DenseNet-121, when optimized, represents a highly accurate model for classifying visual field defects.