Modification of U-Net Architecture Based on Gabor Filters with Trainable Parameters to Improve Biomarker Segmentation on Oct Images

Optical coherence tomography (OCT) is widely used to diagnose retinal pathologies, but the low contrast and noise of OCT images often make automatic biomarker segmentation difficult. This paper aims to improve the accuracy of automatic segmentation of OCT biomarkers by modifying the U-Net architecture. A U-Net-based Gabor-U-Net architecture is proposed that integrates trainable Gabor filters into convolutional layers to improve the recognition of oriented textures and boundaries. Experiments were performed on the OIMHS dataset. The segmentation quality was evaluated by IoU metrics and Dice coefficient. The results show that Gabor-U-Net significantly outperforms classical U-Net in terms of metrics: IoU=0.90 and Dice=0.94 for MH segmentation; IoU <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$=0.88$</tex> and Dice <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$=0.91$</tex> for IRC. This is <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3-4 {\%}$</tex> higher than the base model, which confirms the effectiveness of the proposed modification for accurate segmentation of diagnostically important biomarkers on OCT-images.