Design And Analysis Of WASD Neural Network And BLS With Their Applications To The Auxiliary Diagnosis Of Flatfoot

Flatfoot is a common foot deformity symptom in children and juveniles.If it is not diagnosed and corrected in time,this symptom will continue into adulthood and become an irreversible disease,which will bring great pain and inconvenience to the normal ac-tivities of patients.Therefore,it is of great significance to check whether the feet have flatfoot symptom in time during the critical period of the development of children and juveniles.With the increasing development of computer hardware and theory,machine learning algorithms and their auxiliary medical diagnosis applications have become two research hotspots.Taking these two aspects as the starting point,this paper first studies two types of feedforward neural networks,namely,weight-and-structure-determination(WASD)neural network and broad learning system(BLS).Then establish the models of these two neural networks and carry out a series of pattern classification studies on a foot dataset to realize the auxiliary diagnosis application of flatfoot in juveniles.Fi-nally,a class of pseudo-inverse update algorithms is proposed,and the classification performance of the modified BLS based on such algorithms is experimentally verified.The specific research content of this paper can be summarized as follows:(1)On the basis of in-depth study of WASD algorithm and theory,a WASD neu-ral network model called the mirroring WASD neural network(MWASDNN)is pro-posed,and its corresponding training algorithm is deduced.Moreover,the proposed model is used to conduct pattern classification experiments on a real-world foot dataset,with the experimental results manifesting a satisfactory performance of the model.Specif-ically,classification accuracy of 84.31%and 85.29%are achieved on the left foot data and right foot data,respectively,which is close to the diagnostic level of human physi-cians,indicating that the proposed MWASDNN model can be used as an auxiliary di-agnosis tool for juvenile flatfoot.(2)Through the research of BP-free training algorithms based on pseudo-inverse,the connection between WASD theory and BLS theory is summarized and construct-ed.Then,the classification performance of BLS model and multiple WASD neural network models are compared via a number of experiments.From the experimental re-sults,the effectiveness of the additional structure and algorithms of the BLS compared with the WASD neural network is analyzed,and the efficiency and the limitation(i.e.,the performance of BLS is very dependent on a good initial structure)of the dynamic update algorithms of the BLS are revealed.Finally,an appropriate BLS model struc-ture is selected to carry out pattern classification experiments on the foot dataset.The experimental results show that the accuracy of BLS is better than that of all the WASD models,which confirms that BLS has the application prospects for auxiliary diagnosis of juvenile flatfoot.In addition,through extended experiments on multiple benchmark datasets,the wide applicability of BLS to various types of datasets and the limitations of WASD neural networks in processing image data are demonstrated.(3)On the basis of the Sherman-Morrison formula and Schur complement,two fast pseudo-inverse update algorithms that can be applied to BLS to increase nodes are derived.These two algorithms adopt completely different algorithmic idea from the original BLS dynamic update algorithms,but achieve the same function of the original BLS dynamic update algorithms.Exactly,a new pseudo-inverse matrix is obtained via a small amount of calculation on the basis of the existing pseudo-inverse matrix,thereby avoiding a complete training cycle after adding nodes.The effectiveness of the proposed algorithms is verified through comparative experiments,which demonstrates that the proposed algorithms can be used as alternative algorithms to the original BLS dynamic update algorithms,which provides a new direction for the research of BLS theory.