Evaluating deep learning models for classifying OCT images with limited data and noisy labels.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning (DL) shows promise for classifying Optical Coherence Tomography (OCT) images to diagnose retinal diseases.
• Clinical application is hindered by variability in retinal abnormalities, image noise, and artifacts.
• Data scarcity and label noise present significant challenges for training robust DL models.

Purpose of the Study:

• To evaluate the performance of various DL architectures (ResNet18, ResNet34, ResNet50, VGG16, InceptionV3) for OCT image classification.
• To assess the impact of data scarcity and label noise on DL model accuracy in retinal pathology detection.
• To determine the optimal training data size and strategies for mitigating mislabeling errors.

Main Methods:

• Fine-tuning five pre-trained DL architectures on a dataset of 5526 OCT images.
• Evaluating model performance on reduced subsets, down to 21 images, to simulate data scarcity.
• Assessing the impact of 10%, 15%, and 20% label noise on classification accuracy.

Main Results:

• All DL architectures achieved >90% accuracy with training sets of 345 or more images.
• InceptionV3 demonstrated the highest accuracy (99%) on the full dataset.
• Decreased sample size and increased label noise significantly reduced classification accuracy and increased variability.
• Mitigating 10-20% label noise required 4-14 times more images to match the performance of 345 correctly labeled images.

Conclusions:

• DL models, particularly InceptionV3, can accurately classify retinal pathologies from OCT images when trained on sufficient data (≥345 images).
• Data scarcity and label noise are critical factors affecting DL performance in OCT analysis.
• Increasing training dataset size is an effective strategy to overcome the negative impact of mislabeling errors and improve diagnostic accuracy.