Glaucoma detection in Latino population through OCT's RNFL thickness map using transfer learning

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Glaucoma is a leading cause of irreversible blindness globally, affecting over 60 million people.
• Early diagnosis is crucial for preventing vision loss, yet many cases remain undiagnosed.
• Deep learning (DL) offers potential for objective glaucoma detection, but research in Latino populations is limited.

Purpose of the Study:

• To investigate the efficacy of transfer learning using MobileNet and Inception V3 models for glaucoma detection.
• To analyze optical coherence tomography (OCT) images of retinal nerve fiber layer (RNFL) thickness maps from Mexican patients.
• To address the gap in DL applications for glaucoma screening in the Latino population.

Main Methods:

• Utilized the IBM Foundational Methodology for Data Science.
• Employed MobileNet and Inception V3 architectures for classifying OCT images into glaucomatous and non-glaucomatous categories.
• Retrained models using transfer learning on RNFL thickness map images from 333 OCT files, with 50 images per class for training and 15 per class for prediction.

Main Results:

• MobileNet achieved 86% accuracy for the left eye and 90% for the right eye.
• Inception V3 demonstrated 90% accuracy for both the left and right eyes.
• Both models showed high precision, recall, and F1 scores, indicating strong performance in glaucoma detection.

Conclusions:

• Transfer learning with MobileNet and Inception V3 models is effective for glaucoma detection using OCT RNFL images.
• Inception V3 slightly outperformed MobileNet in classifying left eye images.
• The models demonstrated high accuracy, suggesting their potential as diagnostic aids for ophthalmologists.