IEEE Access · 2023 · 18 citations · 44 references
ComputingComputer ScienceOphthalmologyBiomedical ImagingComputer VisionDiagnostic ImagingDeep LearningDiabetic RetinopathyMedical Image AnalysisRetinal ImagingAutomatic SegmentationConvolutionConvolutional Neural Network SegmentationConvolutional Neural NetworksDiabetic Retinopathy DetectionEncoder-decoder ArchitecturesFeature FusionRetinal Vessel Segmentation
Retinal fundus images contain highly informative geometrical features for detecting diabetic retinopathy (DR), including vessels, especially thin and low-contrast vessels, which are predominant features for accurately diagnosing diabetic retinopathy. Automatic segmentation methods have been developed based on deep convolutional neural networks to replace manual labeling. These methods have shown acceptable performance in fundus vessel segmentation. The UNet model is a well-known architecture of deep neural networks often used for vessel segmentation tasks and has achieved significant performance. However, segmentation tasks remain challenging due to multiple convolutions, down-sampling operations, and inadequate feature fusion in the encoder-decoder architecture. Also, traditional convolution increases the number of multiplications while performing convolution operations. These challenges lead to the loss of information related to thin and low-contrast vessels, eventually affecting the segmentation performance. To tackle this issue, we propose incorporating depthwise parallel attention in the existing UNet framework (DPA-UNet) to achieve accurate vessel segmentation. This approach entails the integration of a depthwise convolution block in the downsampling path and a parallel attention mechanism in the upsampling path of UNet. The primary benefit of depthwise convolution and global information embedding (GIE) is the ability to capture intricate information characteristics across channels. This helps to minimize the information degradation caused by conventional convolution and downsampling techniques. A parallel attention network is proposed in the upsampling path of the existing UNet to optimize the channel and spatial information acquired from the encoder-decoder. Extensive experiments are conducted on three publicly available datasets, namely DRIVE, STARE, and CHASE_DB1, to validate the performance of the proposed model. The findings indicate that the UNET model with depthwise parallel attention achieved a competitive performance with fewer network parameters in segmenting retinal vessels.
44
Deep Residual Learning for Image Recognition
Kaiming He, Xiangyu Zhang, Shaoqing Ren et al. · 2016 · 222.1K citations · Full text
Xception: Deep Learning with Depthwise Separable Convolutions
François Chollet · 2017 · 19K citations
Rethinking Atrous Convolution for Semantic Image Segmentation
Liang-Chieh Chen, George Papandreou, Florian Schroff et al. · arXiv (Cornell University) · 2017 · 7.5K citations · Full text
Dual Attention Network for Scene Segmentation
Jun Fu, Jing Liu, Haijie Tian et al. · 2019 · 6.8K citations
Ridge-Based Vessel Segmentation in Color Images of the Retina
Joes Staal, Michael D. Abràmoff, Meindert Niemeijer et al. · IEEE Transactions on Medical Imaging · 2004 · 4.1K citations