OCT²Former: A retinal OCT-angiography vessel segmentation transformer.

Area of Science:

• Ophthalmology
• Medical Imaging
• Artificial Intelligence

Background:

• Retinal vessel segmentation is crucial for diagnosing eye diseases.
• Segmenting thin vessels and maintaining connectivity are key challenges.
• Optical coherence tomography angiography (OCTA) offers high-resolution retinal imaging.

Purpose of the Study:

• To propose a novel end-to-end transformer-based network, OCT²Former, for accurate retinal vessel segmentation in OCTA images.
• To leverage OCTA's high-resolution characteristics for improved segmentation.
• To address the challenges of segmenting thin vessels and preserving connectivity.

Main Methods:

• An encoder-decoder architecture named OCT²Former was developed.
• It features a dynamic transformer encoder with multi-head dynamic token aggregation attention and an auxiliary convolution branch.
• A lightweight, convolution-based decoder was employed for efficient feature decoding and complexity reduction.

Main Results:

• OCT²Former was evaluated on OCTA-SS, ROSE-1, and OCTA-500 datasets.
• It achieved Jaccard indexes of 0.8344, 0.7855, 0.8099, and 0.8513.
• Performance surpassed the best convolution-based networks by 1.43%, 1.32%, 0.75%, and 1.46% respectively.

Conclusions:

• The proposed OCT²Former demonstrates competitive performance in retinal OCTA vessel segmentation.
• The network effectively segments retinal vessels, addressing key segmentation challenges.
• Results indicate its potential for automatic retinal disease screening and diagnosis.