Research On Retinal Blood Vessel Segmentation Method Based On U-Ne

Retinal vessel segmentation is an important branch of medical image processing that aims to automatically segment the morphological structure of blood vessels in fundus images by computer systems.This technique has a wide range of applications in clinical medicine,such as for early screening of diabetic retinopathy and evaluation before cataract surgery.However,the task remains challenging due to the complex retinal vascular structure and the interference of low contrast,uneven illumination,and pathological exudates in fundus images.In addition,the low contrast of blood vessels is further compounded by the difficulty of accurate segmentation for tiny blood vessels with small diameters,which have less difference in color from the surrounding tissues.With the great success of deep learning techniques for this task in recent years,the mainstream framework U-Net and its variant approaches for this task still have some shortcomings,including the failure to consider global semantic dependencies,the semantic divide between encoder and decoder,and the difficulty of segmenting fine blood vessels.The following studies have been conducted to address the above issues:(1)Intra-and inter-scale augmented-based U-Net segmentation method.U-Netbased methods usually utilize convolutional operations to extract features,however,convolutional operations are a local perceptual model that can only perform feature extraction in a local region of fixed size,lacking consideration of global information,and in addition,simple skip connection operations lack cross-scale interactions and semantic gaps.To this end,a U-Net segmentation method based on Intra-and inter-scale augmentation is proposed,the model is designed from two main perspectives: for the intra-scale encoding-decoding layer,a spatially enhanced self-attentive mechanism is designed to enhance the global and spatial perception during encoding-decoding.Specifically,this mechanism is embedded in each encoding layer to enhance its global spatial aggregation capability,and further extended to the decoding side to alleviate the information loss caused by up-sampling operations during the decoding process.For inter-scale encoding-decoding layer,a novel Inter-scale Fusion module is introduced to enhances semantic interaction with other scales by dynamically selecting rich feature information from the deepest layer.This module further bridges the semantic gap between encoder and decoder,and experimental validation is conducted on three publicly available datasets-DRIVE,CHASE＿DB1,and STARE.The results of the experiments show that the proposed method effectively segments retinal vascular structures and achieves higher improvements in all metrics compared to the baseline model.(2)Cascaded U-Net-based retinal vessel segmentation method.To address the problem of difficult segmentation of fine vessels,a U-Net-based cascade strategy is proposed in this paper to segment vessels from coarse to fine in a two-stage manner to further improve the model’s ability to accurately segment fine vessels,The cascade network consists of three main components: a coarse segmentation model,a calibration module and a fine segmentation model.Among them,the coarse segmentation model is designed as a lightweight U-Net with a self-attention mechanism introduced for extracting regions of interest.Then,the calibration module is introduced to achieve recalibration of the coarse segmentation results by integrating them with the original input information.Finally,the results are used as the input of the fine segmentation model to induce the fine model to focus on significant regions and achieve refinement of the blood vessels.The fine model follows the Intra-and inter-scale augmented U-Net described above.The proposed method has been evaluated on three publicly available datasets(DRIVE,CHASE＿DB1and STARE),and it has been found to improve the segmentation performance compared to the simple cascade approach.