Validation of Deep Learning-Based Automatic Retinal Layer Segmentation Algorithms for Age-Related Macular...

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal layer segmentation in spectral-domain OCT (SD-OCT) is vital for diagnosing and monitoring retinal diseases, including age-related macular degeneration (AMD).
• Deep learning models require expert-annotated ground truth data for training, a process that is time-consuming and limits widespread algorithm adoption across various OCT devices.
• This study addresses the challenge of applying deep learning segmentation models to diverse OCT devices.

Purpose of the Study:

• To validate the robustness of deep learning image segmentation models across multiple OCT devices.
• To assess the ability of trained models to generate clinically relevant metrics for retinal pathologies.
• To evaluate the device independence of state-of-the-art segmentation algorithms.

Main Methods:

• Trained two segmentation algorithms (UNet and DeepLabv3) using SD-OCT images from a Heidelberg-Spectralis device.
• Tested the trained models on SD-OCT images acquired from both Heidelberg-Spectralis and Zeiss-Cirrus devices.
• Evaluated performance using metrics such as mean squared error, mean absolute error (MAE), and Dice coefficients on a dataset spanning healthy to advanced AMD conditions.

Main Results:

• Segmentation models trained on one OCT device (Spectralis) achieved clinically useful results when applied to data from another device (Cirrus).
• Mean Absolute Error (MAE) for internal limiting membrane (ILM) segmentation was 7.0 ± 0.9 μm, and for retinal pigment epithelium (RPE) was 9.5 ± 2.6 μm.
• The Dice similarity coefficient for the RPE drusen complex region reached 0.87 ± 0.01, indicating high agreement between predicted and ground truth segmentations.

Conclusions:

• Segmentation networks can learn domain-independent features from large datasets, allowing application in areas with limited ground truth data.
• The validated deep learning models demonstrate potential for reliable retinal layer segmentation across different OCT devices.
• This approach facilitates broader application of AI in ophthalmology for disease analysis and management.