Customizing CNNs for Blood Vessel Segmentation From Fundus Images Sunil Kumar Vengalil * , Neelam Sinha † , Srinivas S S Kruthiventi ‡ and R Venkatesh Babu § *† International Institute of Information Technology, Bangalore, India ‡§ Indian Institute of Science, Bangalore, India Email: * sunilkumar.vengalil@iiitb.org, † neelam.sinha@iiitb.ac.in, ‡ kssaisrinivas@gmail.com, § venky@cds.iisc.ac.in Abstract—For automatic screening of eye diseases, it is very important to segment regions corresponding to the different eye- parts from the fundal images. A challenging task, in this context, is to segment the network of blood vessels. The blood vessel network runs all along the fundal image, varying in density and fineness of structure. Besides, changes in illumination, color and pathology also add to the difficulties in blood vessel segmentation. In this paper, we propose segmentation of blood vessels from fundal images in the deep learning framework, without any pre-processing. A deep convolutional network, consisting of 8 convolutional layers and 3 pooling layers in between, is used to achieve the segmentation. In this work, a Convolutional Neural Network currently in use for semantic image segmentation is customized for blood vessel segmentation by replacing the output layer with a convolutional layer of kernel size 1 × 1 which generates the final segmented image. The output of CNN is a gray scale image and is binarized by thresholding. The proposed method is applied on 2 publicly available databases DRIVE and HRF (capturing diversity in image resolution), consisting of healthy and diseased fundal images boosted by mirror versions of the originals. The method results in an accuracy of 93.94% and yields 0.894 as area under the ROC curve on the test data comprising of randomly selected 23 images from HRF dataset. The promising results illustrate generalizability of the proposed approach. I. I NTRODUCTION In order to diagnose several diseases of the eye such as Diabetic Retinopathy, Glaucoma, fundal images are captured for examination. Fundal images reveal the important eye re- gions such as optic disk, macula, fovea, blood vessel networks, besides pathologies such as exudates, haemorrages and aneury- sems. Hence examining the fundal images can lead to accurate diagnosis of the disease, if present. However, automating the task of examination, involves several challenges, foremost of which is to separate each of the components into separate regions. One of the difficult components to segment is the blood vessel network that is present almost all over the fundal image, with varying density and fineness of structure. The network around the optic disk is generally denser composed of thicker structures. The vessels branch out along the top and bottom of the optic disk gradually thinning out as they traverse the opposite end of the fundal image. The problem of blood vessel segmentation in fundal images has been well-studied over the past couple of decades, by several groups. Most of the existing techniques utilize case- specific image processing tools [4] [5], and are often hard to generalize due to changes in the acquisition schemes, illumination, color, pathology and image quality. Of late, the framework of ”Deep Learning” has been popular for several computer vision tasks such as object detection and segmen- tation, on natural images. However, for medical applications, the networks need to be customized, to adapt to the specific imaging modality-based issues such as SNR, structural details and spatial resolution. A recent work by Melinak et. al. [1] had explored the utility of Deep learning framework for blood vessel segmentation in fundal images, for a specific dataset DRIVE. In this paper, to incorporate diversity in image resolution and pathology we propose customization and fine- tuning of a CNN architecture [7] to segment the blood vessel network across databases. The rest of the paper is organized as follows. Section II provides an overview of existing techniques for blood vessel segmentation. Section III gives details of our approach includ- ing the CNN architecture used. Results of our experiments on various datasets are given in section IV and section V gives an analysis of the results obtained. Finally section VI concludes the paper. II. RELATED WORK Several methods have been reported to accomplish blood vessel segmentation in fundal images [2], [3]. In [4], the authors represent each pixel by feature vectors comprising of the pixel intensity and two-dimensional Gabor wavelet trans- form responses taken at multiple scales, to extract the blood vessels. Although AUC results of 96% on DRIVE database has been reported using this method, the pre-processing involves altering pixel intensities determined by the camera aperture and choice of ROI. Such issues in preprocessing may lead to loss in the generalizability of the algorithm. In [5], the authors propose use of local and global vessel features cooperatively to segment the vessel network. However, in this work several assumptions are made. The Gaussian function is assumed to be a template for a blood vessel profile. Assumptions are made of fixed width and directions. In [6] the authors carry out scale- space analysis of the first and second derivative of the intensity image. This analysis is claimed to lead to understanding of the 978-1-5090-1746-1/16/$31.00 c  2016 IEEE