A single-step regression method based on transformer for retinal layer segmentation.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal layer segmentation is crucial for diagnosing various eye diseases.
• Current two-step methods and deep learning approaches have limitations in accuracy and data requirements.
• Accurate segmentation of retinal layers from optical coherence tomography (OCT) B-scans is essential for clinical diagnosis.

Purpose of the Study:

• To propose a novel single-step transformer-based method for retinal layer segmentation.
• To improve the accuracy and efficiency of retinal layer segmentation using axial scans (A-Scans).
• To reduce the dependency on large amounts of labeled data for training segmentation models.

Main Methods:

• A single-step transformer-based model was developed for retinal layer segmentation.
• The model was trained using axial data (A-Scans) to identify layer boundaries.
• The method was evaluated on two public datasets: one for diabetic macular edema and another for healthy controls and multiple sclerosis patients.

Main Results:

• The proposed method achieved low average distance errors (0.99 and 3.67 pixels) on the evaluated datasets.
• High accuracy was maintained even with a significant reduction in training data (down to 0.3).
• The method demonstrated state-of-the-art performance, preserving topological correctness.

Conclusions:

• The transformer-based single-step method offers a highly accurate and efficient solution for retinal layer segmentation.
• This approach significantly reduces the need for extensive labeled data, making it more practical for clinical applications.
• The method shows promise for improving the diagnosis of ophthalmological diseases through precise retinal layer analysis.