Developing a 10-Layer Retinal Segmentation for MacTel Using Semi-Supervised Learning.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated segmentation software for optical coherence tomography (OCT) is typically developed and tested on common diseases, leading to limited accuracy in rare pathologies.
• Accurate segmentation of retinal layers and features in OCT is crucial for diagnosing and monitoring eye conditions.

Purpose of the Study:

• To develop a semisupervised deep learning model for segmenting 10 retinal layers and 4 features in OCT images of eyes with Macular Telangiectasia Type II (MacTel).
• To leverage a small labeled dataset and abundant unlabeled images to improve segmentation accuracy for this rare disease.
• To compare the developed model against popular supervised and semisupervised models.

Main Methods:

• Developed a semisupervised deep learning segmentation model.
• Trained the model on a small labeled dataset of MacTel OCT images, augmented by a larger set of unlabeled images.
• Evaluated model performance using intersection over union (IoU) and compared it with existing supervised and semisupervised models.
• Conducted ablation studies to assess the impact of unlabeled data on model performance.

Main Results:

• The semisupervised model significantly outperformed all other tested models in IoU for 10 retinal layers and 2 specific retinal features.
• Performance for segmenting the pre-retinal space and background was similar across all models.
• Increasing the number of unlabeled images used in training improved the performance of the semisupervised model.

Conclusions:

• Leveraging unlabeled data with a semisupervised approach enhances OCT segmentation performance compared to purely supervised methods, especially for rare diseases like MacTel.
• This strategy holds potential for improving segmentation in other conditions with limited labeled data but abundant unlabeled OCT images.