Modified Weights-and-Structure-Determination Neural Network for Pattern Classification of Flatfoot

Abstract

Flatfoot is a common disease in children and juveniles. If the disease is not controlled and treated in time, it may last into adulthood, which can bring a great deal of inconvenience and even pain to daily life. In addition to the diagnosis of the disease simply by doctors and medical equipment, artificial intelligence has become a very promising auxiliary diagnostic tool. In this paper, a neural network with a simple structure is used to classify the foot data to achieve the function of diagnosing flatfoot. The presented neural network is termed as modified weights-and-structure-determination neural network (MWASDNN), of which the input weights are analytically determined by the pseudo-inverse method, while the output weights are randomly generated within a specified interval, and the number of hidden-layer neurons is determined by an incremental method. In addition, the stratified cross-validation method is introduced to choose the model structure that best fits the features of the data set, thereby improving the generalization performance and robustness of the MWASDNN. Utilizing the MWASDNN models to classify the foot data we collected, we finally get the accuracy of 84.31% and 85.29% on the left and right foot data, respectively. Besides, MWASDNN achieves the highest classification accuracy on our foot data set compared to some traditional neural networks, pattern classification methods, and two improved neural networks. These excellent results indicate that the MWASDNN is expected to be designed as a practical flatfoot diagnostic tool.

Keywords

• Flatfoot diagnosis
• modified weights-and-structure-determination neural network (MWASDNN)
• pattern classification
• stratified cross-validation