Automated laser mark segmentation from colored retinal images

Abstract

Medical Image Analysis is an ongoing field of research nowadays. Diabetic Retinopathy (DR) is one of the major diseases being workaround using different techniques of image analysis. Advanced stage of DR is commonly treated with laser at the present time which is a major tool to safe further vision loss which leaves marks on the surface of the retina. We present an automated system for detection of laser marks from colored retinal images to facilitate automated diagnosis of retinal diseases. The proposed system performs preprocessing on the image in order to extract all possible candidate laser mark regions. This is followed by a post processing stage to remove false pixels from candidate regions. The method extracts a number of features for proper representation of all candidate regions. These extracted features are used to facilitate the later classification stage for accurate detection of laser marks from all candidate regions. The validity of a proposed system is performed on a locally gathered retinal image database and results show the significance of proposed system.

Laser Marks Classification for Retinal Images based on Convolutional Neural Network

“…In this study, the system uses 80% of 730 input images for each iteration to train the network and the rest was used for testing it. The highest accuracy was determined by choosing different value of "MaxEpoch" which are (1,2,5,10,20,30,40,50). The best value was selected of the average after conducting five experiments for each values as shown in Table I.…”

“…Dias et al [3] describe a system which is upgrading of a previously introduced an algorithm of retinal quality assessment [4] for detecting the presences of laser marks into digital images of fundus. In [5], Syed et al introduced a method which on classification using support vector machines (SVM). The input is consisting for features of three color domain, two texture domain features and four shape features.…”

“…Layers of Convolutional and batch normalization are typically followed by a nonlinear activation function like a "rectified linear unit" particular by a (ReLU) layer. This layer is used to perform threshold operations for every element, for every input value is negative is equalized to zero, as: (5) This layer does not modify the size of its inputs. Standard ReLU layer has a number of extensions that perform different operations and could improve the performances of some applications.…”

Improving Laser Mark Detection for Retinal Images based on the AlexNet Model

“…The output size in the map of features (number of total neurons) is the product of the height and width. The output size is calculated by: Map Size× Filters Number (4) This type of deep networks may contain single or multi convolutional layers based on the system type, where the data quantity and complexity are the main reason for using multi convolutional layers.…”

“…The size of this area is defined via Max Pooling layer arguments by the pooling size. For example, if the size of pool is (5,4), then the output will return the highest value in the area of height and width of 5 and 4. Also, for the Average Pooling layer, it follows the same process where the output is extracted using average values for the rectangular area for the input of the layer.…”