D2CBDAMAttUnet: Dual-Decoder Convolution Block Dual Attention Unet for Accurate Retinal Vessel Segmentation From Fundus Images

Abstract

Gathering detailed morphological information from retinal blood vessels is crucial in clinical diagnostics, enabling doctors to make precise assessments of patient conditions and to devise custom treatments. Traditional methods of segmenting these vessels from fundus images are not only tedious but require a high degree of specialized knowledge. In light of this, Deep Convolutional Neural Networks (DCNNs), particularly those based on the U-Net architecture, have been acknowledged for their effectiveness in capturing and utilizing contextual features within this context. However, these methods often grapple with challenges such as loss of vital information during pooling and insufficient handling of local context in skip connections, leading to less than optimal results. To address these limitations, this research propounds the novel Convolution Block Dual Attention Module (CBDAM), as well as two pioneering network architectures: The Convolution Block Dual Attention Module Unet (CBDAMUNet) and the Dual Decoder Convolution Block Attention Module with Attention U-Net (D2CBDAMAttUnet). These are built upon a fortified encoder-decoder structure aimed at providing an automated, streamlined detection mechanism from fundus imagery. Thoroughly tested against the DRIVE, CHASEDB1 and STARE datasets, recognized standards in retinal vessel segmentation, the proposed models not only redefine the accuracy benchmarks but also represent a significant stride in automated retinal vessel analysis. The introduction of these advanced networks is a milestone in ophthalmologic diagnostics and research, offering a potent asset for medical professionals and specialists in the field.