Premium
Applicability of deep neural networks and local binary pattern for the classification of MCI‐C and MCI‐NC
Author(s) -
Bhasin Harsh,
Deshwal Vishal
Publication year - 2021
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.053306
Subject(s) - softmax function , dropout (neural networks) , local binary patterns , artificial intelligence , pattern recognition (psychology) , overfitting , support vector machine , computer science , artificial neural network , deep learning , neuroimaging , alzheimer's disease neuroimaging initiative , feature (linguistics) , learning vector quantization , histogram , cognitive impairment , machine learning , cognition , psychology , neuroscience , linguistics , philosophy , image (mathematics)
Background The classification of Mild Cognitive Impairment (MCI) patients, who convert to Alzheimer’s (MCI‐C) and do not convert to Alzheimer’s (MCI‐NC) is important to understand the progression of the disease. The imaging data, particularly s‐MRI helps to understand the disease by capturing the variation in the gray matter. Method We developed a Deep Learning based method to classify MCI‐C and MCI‐NC. For each MRI volume Local Binary Pattern (LBP) was applied on each slice and the features so obtained were concatenated to get the feature vector of the volume. The features were obtained using three methods: creating a histogram having 256 bins, by using the Uniform LBP, and using the Uniform‐Rotation Invariant LBP. This feature set was fed to a Deep Neural Network having two hidden layers and a Softmax layer. Experiments were conducted to find the optimal architecture of the DNN, which gives the best accuracy. Also, a dropout layer was added to prevent overfitting. The results were also compared with the Support Vector Machine. The data used in the experiment was obtained from Alzheimer’s Disease Neuroimaging Initiative (ADNI). Result We evaluated the model performance using accuracy. It was found that amongst the three variants of LBP, that having 256 bins produced the best results. Also, DNN performed far better as compared to SVM. Amongst the various DNN’s the one having 100 units in the first hidden layer, 50 in the next, and having a dropout of 0.1, gave the best results. The model used Adam optimizer and categorical cross‐entropy. It was also noted that the results obtained using raw volumes were inferior as compared to that obtained using LBP. Table 1 shows the results. Conclusion We propose a Deep Learning model for classifying MCI‐C and MCI‐NC. The proposed method is capable of finding discriminating features to classify the two sets. It may be the case that LBP is able to infer the difference in the distribution of microstructures in the two classes which helped the DNN to obtain better accuracy as compared to the state of the art.