Precise segmentation of retinal vessels from fundus images is essential for intervention in numerous diseases, and helpful in preventing and treating blindness. Deep convolutional neural network (DCNN) based approaches have achieved an excellent success in the automatic segmentation of retinal vessels. However, a single convolutional neural network (CNN) structure can only capture limited local features and lack the ability to extract global contexts. Meanwhile, the strategies used for the feature fusion of low-level detail information with high-level semantic information fail to handle the phenomenon of the semantic gap issue between encoder and decoder validly. Therefore, high-precision segmentation of retinal vessels still remains a challenging task. In this paper, a dual-path progressive fusion network, named DPF-Net, is proposed for accurate and end-to-end segmentation of retinal vessels from fundus images. To detect rich feature formation, a dual-path encoder is proposed for effective feature representation, which contains a CNN path for detecting local features and a recurrent convolutional path for extracting contextual information. It could acquire sufficient detailed information and rich contextual information at the same time. In addition, a progressive fusion strategy is proposed for effective feature aggregation at the same scale, adjacent scales and all scales, which is composed by interactive fusion (IF) block, cross-layer fusion (CLF) block and a scale feature fusion (SFF) block. Combine with the feature maps from different paths at the same scale, an IF block is proposed to fuse detailed features with contextual features to obtain fusion features. Meanwhile, a CLF block is proposed to fuse features between adjacent scales to guide low-level feature representation through high-level features. Finally, a SFF block is proposed to recalculate the weights of all scales to realize effective feature aggregation from all scales. Extensive experiments have conducted on three publicly available retinal datasets (DRIVE, CHASEDB1 and STARE). Experimental results show that proposed DPF-Net could achieve a better segmentation results compared to other state-of-the-art methods, especially the proposed progressive fusion strategy indeed promotes feature fusion and significantly boosts the segmentation performance.
The deterioration of numerous eye diseases is highly related to the fundus retinal structures, so the automatic retinal vessel segmentation serves as an essential stage for efficient detection of eye-related lesions in clinical practice. Segmentation methods based on encode-decode structures exhibit great potential in retinal vessel segmentation tasks, but have limited feature representation ability. In addition, they don’t effectively consider the information at multiple scales when performing feature fusion, resulting in low fusion efficiency. In this paper, a newly model, named DEF-Net, is designed to segment retinal vessels automatically, which consists of a dual-encoder unit and a decoder unit. Fused with recurrent network and convolution network, a dual-encoder unit is proposed, which builds a convolutional network branch to extract detailed features and a recurrent network branch to accumulate contextual features, and it could obtain richer features compared to the single convolution network structure. Furthermore, to exploit the useful information at multiple scales, a multi-scale fusion block used for facilitating feature fusion efficiency is designed. Extensive experiments have been undertaken to demonstrate the segmentation performance of our proposed DEF-Net.
26 November 2024 in Biomedical Signal Processing and Control