Effectiveness Of Deep Learned Features In Categorizing Heterogenous Neurological Disorders

Neurological disorders are found to exist commonly among elderly and young adults. Studies states that these disorders pose severe threat to the individual for performing his day living activities. This increases the need to identify the type of disease at an early stage. In this study, 165 MR images comprising Alzheimer disease, Alzheimer visual agnosia, Cerebral disease, Huntington disease, Motor neuron disease, Pick’s disease obtained from AANLIB database are considered. Squeezenet and Inception V3 deep learning models are used to extract pixel level features. Efficiency of logistic regression in classifying the abnormalities is accessed by varing the regularizer. Results show that the features extracted using squeezenet and inception V3 models has the ability of differentiating the disorders. Logistic regression with ridge and lasso regularizer fails to identify the abnormalities using features of inception V3 model. Logistic regression without regularizer has an accuracy of 75.8% and 0.695 of MCC indicates that logistic regression exhibit good predictabilty of actual classes. Meanwhile, the ridge regularizer based classificaation using squeezenet features has comparably better performance with 84.2% accuracy over 78.2% accuracy of logistic regression without regularizer. From the study it is concluded that the squeezenet features has better ability than the features obtained from inception V3 model. This may be due to the overfitting of data which results as a results of increased feature length of inception model compared to squeezenet model.