A new deep learning method for blood vessel segmentation in retinal images based on convolutional kernels and modified U-Net model.

BACKGROUND AND OBJECTIVE: Automatic monitoring of retinal blood vessels proves very useful for the clinical assessment of ocular vascular anomalies or retinopathies. This paper presents an efficient and accurate deep learning-based method for vessel segmentation in eye fundus images. METHODS: The approach consists of a convolutional neural network based on a simplified version of the U-Net architecture that combines residual blocks and batch normalization in the up- and downscaling phases. The network receives patches extracted from the original image as input and is trained with a novel loss function that considers the distance of each pixel to the vascular tree. At its output, it generates the probability of each pixel of the input patch belonging to the vascular structure. The application of the network to the patches in which a retinal image can be divided allows obtaining the pixel-wise probability map of the complete image. This probability map is then binarized with a certain threshold to generate the blood vessel segmentation provided by the method. RESULTS: The method has been developed and evaluated in the DRIVE, STARE and CHASE_Db1 databases, which offer a manual segmentation of the vascular tree by each of its images. Using this set of images as ground truth, the accuracy of the vessel segmentations obtained for an operating point proposal (established by a single threshold value for each database) was quantified. The overall performance was measured using the area of its receiver operating characteristic curve. The method demonstrated robustness in the face of the variability of the fundus images of diverse origin, being capable of working with the highest level of accuracy in the entire set of possible points of operation, compared to those provided by the most accurate methods found in literature. CONCLUSIONS: The analysis of results concludes that the proposed method reaches better performance than the rest of state-of-art methods and can be considered the most promising for integration into a real tool for vascular structure segmentation.

A tool for automated diabetic retinopathy pre-screening based on retinal image computer analysis.

AIM: This paper presents a methodology and first results of an automatic detection system of first signs of Diabetic Retinopathy (DR) in fundus images, developed for the Health Ministry of the Andalusian Regional Government (Spain). MATERIAL AND METHODS: The system detects the presence of microaneurysms and haemorrhages in retinography by means of techniques of digital image processing and supervised classification. Evaluation was conducted on 1058 images of 529 diabetic patients at risk of presenting evidence of DR (an image of each eye is provided). To this end, a ground-truth diagnosis was created based on gradations performed by 3 independent ophthalmology specialists. RESULTS: The comparison between the diagnosis provided by the system and the reference clinical diagnosis shows that the system can work at a level of sensitivity that is similar to that achieved by experts (0.9380 sensitivity per patient against 0.9416 sensitivity of several specialists). False negatives have proven to be mild cases. Moreover, while the specificity of the system is significantly lower than that of human graders (0.5098), it is high enough to screen more than half of the patients unaffected by the disease. CONCLUSION: Results are promising in integrating this system in DR screening programmes. At an early stage, the system could act as a pre-screening system, by screening healthy patients (with no obvious signs of DR) and identifying only those presenting signs of the disease.

Obtaining optic disc center and pixel region by automatic thresholding methods on morphologically processed fundus images.

Development of automatic retinal disease diagnosis systems based on retinal image computer analysis can provide remarkably quicker screening programs for early detection. Such systems are mainly focused on the detection of the earliest ophthalmic signs of illness and require previous identification of fundal landmark features such as optic disc (OD), fovea or blood vessels. A methodology for accurate center-position location and OD retinal region segmentation on digital fundus images is presented in this paper. The methodology performs a set of iterative opening-closing morphological operations on the original retinography intensity channel to produce a bright region-enhanced image. Taking blood vessel confluence at the OD into account, a 2-step automatic thresholding procedure is then applied to obtain a reduced region of interest, where the center and the OD pixel region are finally obtained by performing the circular Hough transform on a set of OD boundary candidates generated through the application of the Prewitt edge detector. The methodology was evaluated on 1200 and 1748 fundus images from the publicly available MESSIDOR and MESSIDOR-2 databases, acquired from diabetic patients and thus being clinical cases of interest within the framework of automated diagnosis of retinal diseases associated to diabetes mellitus. This methodology proved highly accurate in OD-center location: average Euclidean distance between the methodology-provided and actual OD-center position was 6.08, 9.22 and 9.72 pixels for retinas of 910, 1380 and 1455 pixels in size, respectively. On the other hand, OD segmentation evaluation was performed in terms of Jaccard and Dice coefficients, as well as the mean average distance between estimated and actual OD boundaries. Comparison with the results reported by other reviewed OD segmentation methodologies shows our proposal renders better overall performance. Its effectiveness and robustness make this proposed automated OD location and segmentation method a suitable tool to be integrated into a complete prescreening system for early diagnosis of retinal diseases.

Locating the fovea center position in digital fundus images using thresholding and feature extraction techniques.

A new methodology for detecting the fovea center position in digital retinal images is presented in this paper. A pixel is firstly searched for within the foveal region according to its known anatomical position relative to the optic disc and vascular tree. Then, this pixel is used to extract a fovea-containing subimage on which thresholding and feature extraction techniques are applied so as to find fovea center. The methodology was evaluated on 1200 fundus images from the publicly available MESSIDOR database, 660 of which present signs of diabetic retinopathy. In 93.92% of these images, the distance between the methodology-provided and actual fovea center position remained below 1/4 of one standard optic disc radius (i.e., 17, 26, and 27 pixels for MESSIDOR retinas of 910, 1380 and 1455 pixels in size, respectively). These results outperform all the reviewed methodologies available in literature. Its effectiveness and robustness with different illness conditions makes this proposal suitable for retinal image computer analyses such as automated screening for early diabetic retinopathy detection.