Diabetic retinopathy, an eye disease commonly afflicting diabetic patients, can result in loss of vision if prompt detection and treatment are not done in the early stages. Once the symptoms are identified, the severity level of the disease needs to be classified for prescribing the right medicine. This study proposes a deep learning-based approach, for the classification and grading of diabetic retinopathy images. The proposed approach uses the feature map of ResNet-50 and passes it to Random Forest for classification. The proposed approach is compared with five state-of-the-art approaches using two category Messidor-2 and five category EyePACS datasets. These two categories on the Messidor-2 dataset include ’No Referable Diabetic Macular Edema Grade (DME)’ and ’Referable DME’ while five categories consist of ‘Proliferative diabetic retinopathy’, ‘Severe’, ‘Moderate’, ‘Mild’, and ‘No diabetic retinopathy’. The results show that the proposed approach outperforms compared approaches and achieves an accuracy of 96% and 75.09% for these datasets, respectively. The proposed approach outperforms six existing state-of-the-art architectures, namely ResNet-50, VGG-19, Inception-v3, MobileNet, Xception, and VGG16.
1 January 1986 in IEEE Transactions on Reliability
A quantitative reliability model for a phased mission system is developed using a Markov process. Two cases for the mission-phase change times are assumed: 1) to be known in advance and 2) to be random variables. A method of solution is presented and illustrated by examples. The solution of phased-mission systems is equivalent to solving a sequence of uni-phase systems with appropriate initial conditions.
1 July 1988 in IEEE Transactions on Automatic Control
Recently, CNNs have become very popular in the machine learning field, due to their high predictive power in classification problems that involve very high dimensional data with tens of hundreds of different classes. CNN is a natural extension to MLP with few modifications which resulted in a breakthrough. Mainly, the MLP algebraic dot product as a similarity function was replaced with 2-d convolution; in addition to a pooling layer which reduces parameter dimensions making the model equi-variant to translations, distortions, and transformations. The sparse connectivity nature of CNN is also a variation to the MLP. The two models were implemented on the EMNIST dataset which was used as 50% and 100% of its capacity. The models were trained with fixed and flexible number of epochs in two runs. Using 100% of EMNIST; for the fixed run CNN achieved test accuracy of 92% and MLP 31.43%, where in the flexible run the CNN achieved 92% and MLP 89.47%. Using 50% of EMNIST; for the fixed run CNN achieved test accuracy of 92.9% and MLP 33.75%, where in the flexible run of 92.9% and MLP 88.20%. The CNN demonstrated a good maintenance of high accuracy for image like inputs and also proved to be a better candidate for big data applications.