Extended U-net for Retinal Vessel Segmentation

Abstract

The retinal vascular tree is an important biomarker for the diagnosis of ocular dis-ease, where an efficient segmentation is highly required. Recently, various standard Convolutional Neural Networks CNN dedicated for segmentation are applied for retinal vessel segmentation. In fact, retinal blood vessels are presented in different retinal image resolutions with a complicated morphology. Thus, it is difficult for the standard configuration of CNN to guarantee an optimal feature extraction and efficient segmentation whatever the image resolution is. In this paper, new retinal vessel segmentation approach based on deep learning architecture is propounded. The idea consists of enlarging the kernel size of convolution layer in order to cover the vessel pixels as well as more neighbors for extracting features. Within this objective, our main contribution consists of identifying the kernel size in correlation with retinal image resolution through an experimental approach. Then, a novel U-net extension is proposed by using convolution layer with the identified kernel size. The suggested method is evaluated on two public databases DRIVE and HRF having different resolutions, where higher segmentation performances are achieved respectively with 5*5 and 7*7 convolution kernel sizes. The average accuracy and sensitivity values for DRIVE and HRF databases are respectively in the order of to 0.9785, 0.8474 and 0.964 and 0.803 which outperform the segmentation performance for the standard U-net.

Retinal blood vessel segmentation from high resolution fundus image using deep learning architecture

“…This significant enhancement is attributed to the integration of multi-scale convolutional kernel sizes into the U-Net model. Additionally, we can conclude that the proposed method's sensitivity rate outperforms DL-based methods such as [8], [9], [10], [3], [1], [5] and [11] exhibiting a notable difference of approximately 14%. Consequently, the proposed network demonstrates a robust capability to detect vessel pixels.…”

“…Additionally, works like [1] and [4] suggest switching standard convolution layers within U-Net blocks with dilated convolution layers parameterized by a 3x3 kernel dilated by a factor of 2 and depthwise convolution modules. Even, several have updated the convolution kernel size respect to vessel representation in fundus images, such as the work described in [5].…”

“…As described in section 2, the approach proposed in [5] and the work [6] involves enlarging the convolution kernel size to capture more surrounding neighbors during computation, improving results on high-resolution fundus images when using a 7x7 kernel size. However, applying a large kernel size throughout the entire network may miss vessel details represented on a few pixels and may mix features from different vessel thicknesses.…”

“…The proposed model is validated using the high-resolution retinal database HRF, comprising 45 retinal images sized at 3504x2336 with a resolution of approximately 6µm/pixel. Various evaluation metrics, including Accuracy (Acc), Sensibility (Sens), Specificity (Spec), and DICE, are employed for pixel classification results calculated as equations (2)(3)(4)(5). The implementation is performed on an Intel Core i7 processor with a frequency of 3.67 GHz, 8GB RAM, and a NVIDIA GTX 980 GPU using CUDA 9.0 with CUDNN 7.6.3.…”

Deep Learning Architecture with an Optimized Convolutional Processing for the Segmentation of Retinal Blood Vessels

“…Fundus imaging blood vessel segmentation methods has been the subject of various reviews, including [7], which examined different categories of approaches. Several wellknown CNNs are proposed in the literature and have been applied for this task, which are characterized by the ability of learning complex features.…”