BEFD: Boundary Enhancement and Feature Denoising for Vessel Segmentation

Abstract

Blood vessel segmentation is crucial for many diagnostic and research applications. In recent years, CNN-based models have leaded to breakthroughs in the task of segmentation, however, such methods usually lose high-frequency information like object boundaries and subtle structures, which are vital to vessel segmentation. To tackle this issue, we propose Boundary Enhancement and Feature Denoising (BEFD) module to facilitate the network ability of extracting boundary information in semantic segmentation, which can be integrated into arbitrary encoderdecoder architecture in an end-to-end way. By introducing Sobel edge detector, the network is able to acquire additional edge prior, thus enhancing boundary in an unsupervised manner for medical image segmentation. In addition, we also utilize a denoising block to reduce the noise hidden in the low-level features. Experimental results on retinal vessel dataset and angiocarpy dataset demonstrate the superior performance of the new BEFD module.

Global Transformer and Dual Local Attention Network via Deep-Shallow Hierarchical Feature Fusion for Retinal Vessel Segmentation

“…The dilated convolution, which expands the receptive area without parameters increasing, is widely adopted in vessel segmentation [39]. Additionally, to prevent the attenuation of the edge information of blood vessels, researchers employed a combination of unsupervised edge detection and deep networks [41]. Therefore, combining the dilated convolution and edge detection method is an effective method to capture the local characterizations of neighboring pixels while expanding the local receptive field of the network.…”

“…To adjust the receptive field more flexibly without increasing computational cost, the dilated convolution was applied to replace the original convolution for medical images segmentation tasks [40]. To address the edge attenuation in segmentation results, Zhang et al also studied a boundary enhancement method with Sobel operators, alleviating the loss of vessel edge [41]. In addition, recent studies showed that the deep features and shallow features in a deep neural network can be complementary [42].…”

Spine Image Segmentation Based on U-Net and Atrous spatial pyramid pooling

“…The shallower high-resolution layer is used to solve the problem of pixel positioning, and the deeper layer is used to solve the problem of pixel classification [5]. However, the U-Net still has some disadvantages and shortcomings, such as：effective convolution increases the difficulty, universality of model design, and the form of trimming is not symmetrical with the Feature Map [6].…”

Study Group Learning: Improving Retinal Vessel Segmentation Trained with Noisy Labels

“…Conventionally, hand-crafted filters [12,13,16,27] like Gabor [13] and Gaussian-based ones [12] are explored to extract features for pixel selection, vessel clustering and segmentation. Recently, data-driven based methods utilize UNet-based model [17] or its variants [24,25,28,29,10] to achieve significant performance compared with traditional methods. Those deep learning methods focus on the design of UNet structures with better feature representation [28,24], or the decouple of structure and textures of retinal images [29].…”

“…Recently, data-driven based methods utilize UNet-based model [17] or its variants [24,25,28,29,10] to achieve significant performance compared with traditional methods. Those deep learning methods focus on the design of UNet structures with better feature representation [28,24], or the decouple of structure and textures of retinal images [29]. However, datadriven methods highly suffer from over-fitting issues when the given training data is insufficient.…”