Classification of Digital Intra Oral Periapical Radiographs Represented as FeatureVector Using Neural Network Architectures: A Comparative Study

Study of various multi-layer perceptron neural network architectures and their comparison for classification of Intra Oral Periapical radiographs (IOPA) where individual tooth in that radiograph is represented by its feature vector. In this paper input vector for the neural network is a seven dimensional feature vector of individual tooth. Data vectors are distributed in ratio 4:1 for training and testing purposes. The data used for testing is redistributed in 3:1 ratio for training and validation. Four different neural network architectures are compared based on their performance for the same data set. Neural networks realized in this paper are perceptron classifier using threshold logic neuron, perceptron classifier using threshold logic neuron with tower algorithm for structure growing, feed forward neural network and back propagation neural network. Multi-layer perceptron neural network architectures are tested on 10 radiographic images containing healthy as well as diseased tooth. 30 images are used for training and 10 images are used for validation. The classification algorithm presented in this research work is a two class classifier. This algorithm can be very easily adopted for multi class classification by calling the same repetitively using the classification strategy one against rest. Architectures used in this paper differ in selection of activation logic for hidden layer neurons and output layer neurons. Input layer in all architectures comprises of linear neurons. Hidden layer neurons are threshold logic neurons or sigmoidal neurons. Output layer neurons are threshold logic neurons. The algorithm can be improved by designing multi-class classifier for the same data.