C.A. Mohamed (Canada), H.S. Hasan (UAE), and G. Auda (Canada)
Intelligent systems, modular neural networks, pattern recognition, diabetic retinopathy, voting schemes.
An automatic intelligent screening system for diabetic retinopathy is proposed. The system uses the pattern recognition capabilities of feedforward neural networks to perform the diagnosis. However, due to the irregularities in the visual shapes of the symptoms and the large number of inputs of the network, a non-modular classifier was not practically good. Using three techniques of image pre-processing, three sets of features were extracted and a Modular Neural Network (MNN) is used to give much better results. Two types of retinopathy symptoms were screen7ed: hemorrhages and exudates. A number of cases were extracted from Fundus photographs and were manually diagnosed by an Ophthalmologist. The dataset was then divided into two subsets, one was used for the training process of the neural networks and the other is used for testing their performance. A voting technique was used to make the final decision of the MNN using the local decisions of its modules. The intelligent system accurate-recognition percentages, when compared to the Ophthalmologist diagnoses, reached 96.75% and 79.75% for the hemorrhages and the exudates symptoms, respectively. The MNN results were higher than the best non-modular networks by 17% and 6%, approximately, for the hemorrhages and the exudates patterns, respectively.
Important Links:
Go Back
