Deep ocular tumor classification model using cuckoo search algorithm and Caputo fractional gradient descent.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Interpreting digital ocular fundus images for tumor diagnosis is challenging due to image complexity and subtle tumor features.
• Automated detection of ocular tumors is vital for prompt diagnosis and effective treatment.

Purpose of the Study:

• To investigate a robust deep learning system for classifying ocular tumors using fundus images.
• To introduce and evaluate a novel optimizer integrating Caputo fractional gradient descent (CFGD) with the cuckoo search algorithm (CSA) for enhanced accuracy and convergence speed.

Main Methods:

• Trained Vgg16, AlexNet, and GoogLeNet models on 400 ocular fundus images (benign vs. malignant) using the proposed CFGD-CSA optimizer.
• Compared the novel optimizer's performance against existing methods like SGDM, ADAM, CSA, CFGD, BASADAM, and CSA-ADAM.
• Evaluated performance based on accuracy, robustness, consistency, and convergence speed.

Main Results:

• The proposed optimizer achieved mean accuracies of 86.43% (Vgg16), 87.42% (AlexNet), and 87.62% (GoogLeNet).
• Demonstrated significant enhancements in classification accuracy, robustness, consistency, and convergence speed compared to existing approaches.
• The novel optimizer shows substantial potential for improving deep learning model performance in medical image classification.

Conclusions:

• The developed deep learning system with the novel optimizer offers a promising approach for accurate ocular tumor identification.
• This research contributes to advancing computer-aided diagnosis systems for ocular tumors, highlighting the optimizer's benefits in medical image analysis.