xiaowuc2 / vae-alzheimer-detection-using-imbalanced-mri-images

Variational Autoencoder based Imbalanced Alzheimer detection using Brain MRI Images

Variational Autoencoder: Alzheimer Detection Using Imbalanced MRI Images / ​Code​/ ​Research Paper​

Abstract

Alzheimer’s disease is an incurable, progressive neurologicaldisease that cause the brain to shrink and brain cells to die. Earlierdetection of Alzheimer’s disease can lead to proper treatment and pre-vent brain tissue damage. So accurate and fast detection of Alzheimer ismost important. There is many well established model which can predictAlzheimer’s disease but they struggle hard for imbalance datasets. Thiswork has foused on the early and fast detection of Alzheimer’s diseasefrom Imbalance Brain Magnetic resonance imaging(MRI) dataset usingproper Variational Autoencoder(VAE) model. Firstly, the whole datasethas been converted into a latent vector form then we used some popu-lar sampling methods to balance the dataset. Finally, we applied someclassification algorithm to classify the Brain MRI images into four dif-ferent classes‘Very Mild Demented’,‘Mild Demented’,‘Non Demented’and ‘Moderate Demented’. In order to maintain the stability in resultwe have applied 10-fold cross validation. In our experiment it indicatesimprovement in results to detect Alzheimer’s disease using VAE.

Proposed Method

Variational Autoencoder

In the case of Autoencoders the latent vector we get from the encoder part onlycontains lattent attributes but not in a probabilistic fashion. That’s why we consider Variational Autoencoder(VAE)[10] for our experiment. VAE provide thelatent vector which contains lattent attributes as a probability distribution. Butin this consent a question may arise why not we are using Principal componentsanalysis(PCA) for the dimensionality reduction. The solution is, in the case ofdimensionality reduction PCA only perform linear dimensionality reduction butVAE performs large-scale non-linear dimensionalilty reduction also VAE canreduce the dimensionality fast and accurately without losing much information.The VAE model the input data as follows:

pθ(x|z) =f(x;z,θ)    p(z) =N(z|0,I)

Fig. 1 Variational Autoencoder is trained and Encoder is later used to obtainlatent vectors.

Latent Vector Resampling Techniques

Till now we have got latent vector by feeding the original imbalance data toour proposed VAE. Fig 2 depicts how our proposed VAE is generating la-tent vector from input image data. In this section we are mainly focusing onthe trainning biasing problem due to the imbalance in the dataset. In orderto resolve this imbalance problem we are applying some popular resamplingtechniques including, oversampling methods (Synthetic Minority OversamplingTechnique (SMOTE)[6], Adaptive Synthetic Sampling (ADASYN)[9], Borde-line SMOTE[8](for our experiment we used BorderlineSMOTE-1) and SVMSMOTE[13]) followed by some undersampling techniques (Cluster Centroid[20]. omek’s Links[16], Edited Nearest Neighbour(ENN)[18], Neighbourhood Clean-ing Rule(NHC Rule)[11] and All KNN[15]) and two hybrid sampling methodviz., SMOTE ENN[2] and SMOTE Tomek[3]

Fig. 2 Proposed Encoder Architecture.

Citation