Deep Learning and Transfer Learning for Optic Disc Laterality Detection: Implications for Machine Learning in...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) achieves expert-level performance in medical image classification, including ophthalmology.
• This study focuses on a DL system for determining optic disc laterality (right vs. left eye).
• The system is designed to function with both normal and abnormal optic discs.

Purpose of the Study:

• To develop and evaluate a DL system for accurate optic disc laterality determination.
• To assess the performance of a modified ResNet-152 deep convolutional neural network (DCNN) for this task.
• To demonstrate the utility of transfer learning in neuro-ophthalmology image analysis.

Main Methods:

• Transfer learning was applied to the ResNet-152 DCNN, pretrained on ImageNet.
• A dataset of 576 color fundus photographs, including normal and abnormal optic discs, was used.
• Performance was evaluated using 5-fold cross-validation and receiver operating characteristic curves (AUC).

Main Results:

• The DCNN achieved an average AUC of 0.999 (±0.002) and accuracy of 98.78% (±1.52%) in cross-validation.
• External validation on a separate dataset yielded an average AUC of 0.996 (±0.005) and accuracy of 97.2% (±2.0%).
• High sensitivity and specificity were consistently achieved in both validation phases.

Conclusions:

• High-performing DL systems can be developed from small datasets for neuro-ophthalmology image labeling.
• The study highlights the power of transfer learning for tasks like optic disc laterality determination.
• The developed DCNN can aid in curating medical image databases and streamlining ophthalmologist workflows.