An Improved Semantic Segmentation Method for Retinal OCT Images Based on High-Resolution Network and Polarized Self-Attention Mechanism

Optical Coherence Tomography (OCT) is an efficient, non-invasive imaging technique. It operates by scanning ocular tissues with a laser beam and reconstructing cross-sectional images of the eye tissues using reflected light signals. Traditional methods for segmenting retinal OCT image structures rely on experience and suffer from poor repeatability and low accuracy. In response to this need, we have developed a novel end-to-end method for segmenting eight layers of tissue structures in retinal OCT images, based on High-Resolution Networks (HRNet), demonstrating outstanding performance. We have enhanced the base model by incorporating a polarized self-attention mechanism and modifying the loss function of the model. This has resulted in significant improvements. Achieved a good score of 0.8355 on the Miou coefficient, and can accurately segment out eight layers of tissue structure in OCT images.