A comparison of deep learning U-Net architectures for posterior segment OCT retinal layer segmentation

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning, particularly U-Net architectures, has advanced automated retinal layer segmentation in Optical Coherence Tomography (OCT) images.
• Previous comparisons of U-Net variants for OCT segmentation lack comprehensiveness, often using mismatched networks or varied datasets, hindering accurate performance assessment.

Purpose of the Study:

• To conduct a detailed, unbiased comparison of eight U-Net architecture variants for retinal layer segmentation across diverse OCT datasets.
• To evaluate the necessity of complex U-Net variants over the standard architecture for OCT retinal layer segmentation.

Main Methods:

• Evaluated eight U-Net architecture variants on four distinct OCT datasets, varying in population, pathology, and acquisition parameters.
• Utilized the Dice coefficient to assess segmentation performance across all tested architectures and datasets.
• Ensured network matching by considering an equivalent number of layers for a fair comparison.

Main Results:

• Minimal performance differences were observed between most U-Net variants and the standard U-Net architecture across all datasets.
• A minor performance improvement was noted with the addition of an extra convolutional layer per pooling block in all architectures.
• The standard U-Net architecture demonstrated sufficient performance for OCT retinal layer segmentation.

Conclusions:

• The vanilla U-Net is adequate for retinal layer segmentation in OCT images, suggesting complex, state-of-the-art variants may be unnecessary.
• Careful architecture matching is crucial for unbiased performance evaluation of deep learning models in medical image segmentation.
• Selecting simpler U-Net models can save significant time and cost in research and clinical practice without compromising segmentation accuracy.