Contents lists available at ScienceDirect Medical Hypotheses journal homepage: www.elsevier.com/locate/mehy DCCMED-Net: Densely connected and concatenated multi Encoder-Decoder CNNs for retinal vessel extraction from fundus images Ümit Budak a, ⁎ , Zafer Cömert b , Musa Çıbuk c , Abdulkadir Şengür d a Department of Electrical and Electronics Engineering, Bitlis Eren University, Bitlis, Turkey b Department of Software Engineering, Samsun University, Samsun, Turkey c Department of Computer Engineering, Bitlis Eren University, Bitlis, Turkey d Department of Electrical and Electronics Engineering, Technology Faculty, Firat University, Elazig, Turkey ABSTRACT Recent studies have shown that convolutional neural networks (CNNs) can be more accurate, efficient and even deeper on their training if they include direct connections from the layers close to the input to those close to the output in order to transfer activation maps. Through this observation, this study introduces a new CNN model, namely Densely Connected and Concatenated Multi Encoder-Decoder (DCCMED) network. DCCMED contains concatenated multi encoder-decoder CNNs and connects certain layers to the corresponding input of the subsequent encoder-decoder block in a feed-forward fashion, for retinal vessel extraction from fundus image. The DCCMED model has assertive aspects such as reducing pixel-vanishing and encouraging features reuse. A patch-based data augmentation strategy is also developed for the training of the proposed DCCMED model that increases the generalization ability of the network. Experiments are carried out on two publicly available datasets, namely Digital Retinal Images for Vessel Extraction (DRIVE) and Structured Analysis of the Retina (STARE). Evaluation criterions such as sensitivity (Se), specificity (Sp), accuracy (Acc), dice and area under the receiver operating characteristic curve (AUC) are used for verifying the effectiveness of the proposed method. The obtained results are compared with several supervised and unsupervised state-of-the-art methods based on AUC scores. The obtained results demonstrate that the proposed DCCMED model yields the best performance compared with the-state-of-the-art methods according to accuracy and AUC scores. Introduction Automatic segmentation of retinal vessels plays a critical role in the computer-aided diagnosis (CAD) of fundus images since this approach supports clinical studies and early diagnoses of diseases such as diabetic retinopathy, arteriosclerosis, age-related macular degeneration, hy- pertension, and glaucoma [1]. The CAD creates an opportunity to ex- amine the morphological structure of retinal vasculature in a quanti- tative manner through a fundus camera as a useful means to providing non-invasively valuable visual clues regarding ophthalmologic diseases [2]. Moreover, the automatic segmentation of vessels is not only a critical operation for the retina, but is also useful in several biomedical applications such as the reconstruction of coronary arteries, hemody- namic analysis of vascular tress, computer-assisted laser surgery, and biometric identification [3]. Although automatic segmentation of ves- sels has great potential in terms of several biomedical applications, providing a robust, consistent, and effective vessel segmentation model is still a considerably difficult task [4]. The field presents some major drawbacks from the limitation of current imaging techniques as fundus images have normally low-contrast regions with high noise levels. Also, manual segmentation is a time-consuming, labor-intensive, and error- prone process that requires rigorous attention even for an experienced eye care specialist. Lastly, there can be significantly discrete areas that exhibit different morphological characteristics in terms of vessel width, direction, curvature, branching patterns, and tortuosity in a single fundus image. For these reasons, automated or semi-automated CAD systems are in demand as they produce quantitative sets of measure- ments based on the clinician’s knowledge, ensuring repeatable results, and affording beneficial functionalities for the identification and sum- marizing of key information in retinal examinations. Furthermore, it is possible to prevent major vision loss by employing these systems and with proper treatment planning. In recent years retinal vessel segmentation has become an attractive subject, with various models proposed to cope with its known dis- advantages. These models can be collated within several basic cate- gories such as pixel-based [5,6] windows-based [7], classifier-based [8–10], and tracking method-based [11–13]. The pixel-based methods take on the issue as a binary classification problem consisting of a vessel region and background. Although these methods are useful in general diagnosis, they involve additional calculations for applications re- quiring structural information. Window-based methods such as edge detection, predict a match at each pixel for a given model against the pixel’s surrounding window [7]. Classifier-based method normally consist of two steps. First, a low-level algorithm is employed to yield a https://doi.org/10.1016/j.mehy.2019.109426 Received 2 August 2019; Accepted 9 October 2019 ⁎ Corresponding author at: Electrical and Electronics Engineering Department, Bitlis Eren University, 13100, Bitlis, Turkey. E-mail address: ubudak@beu.edu.tr (Ü. Budak). Medical Hypotheses 134 (2020) 109426 0306-9877/ © 2019 Elsevier Ltd. All rights reserved. T