P2‐331: A HYBRID MACHINE LEARNING UNSUPERVISED APPROACH FOR EARLY DETECTION OF NEURODEGENERATIVE ALZHEIMER'S DISEASE

Abstract

The heterogeneous information of MRI needs to be processed in order to extract information exploitable within the context of medical perceptive. The proposed technique is one of the best approach for early diagnosis identifying homogeneous regions in brain MRI. We have introduced the hybridization of two different deep learning method (supervised and unsupervised)with their unique feature of mapping and training with variation in neuron position (Adaptive moving self-organizing map) and efficient clustering (k-means)for tissue extraction. The OASIS data with Group 1: 95 subjects aged 62-86years, mild AD (CDR = 1): 25 AD, 64 NC, Group 2: 148 subjects aged 67–86 years, mild AD (CDR = 1): 25 AD,123 NC Group 3: 53 subjects aged 76–96 years, mild and very mild AD (CDR = {1,0.5}): 17AD, 36 NC. Group 4: 194 subjects aged 62–96 years, moderate, mild and very mild AD (CDR = {2,1,0.5}): 100 AD, 94 NC is used. The process starts with skull removing and noise removal using threshold and contrast enhancement as pre-processing shown in Figure 1. After that 24 features are extracted and is used as criteria for clustering and mapping with adaptive moving K- means. The set of OASIS MRI volume data with ground truth of FSL segmentation is used for comparison with proposed algorithm. The results produced for classifying GM, WM and CSF is shown in Figure 2-5. The result produced extracted ventricle and hippocampal region from total brain volume (axial, coronal and sagittal) indicating early biomarkers AD shown in Figure 6. The neurological tests determines the base of accuracy and facts are actually criteria of classification assessed by calculating the accuracy rate. It includes accuracy (Acc), sensitivity (Sens), specificity (Spec), Similarity criteria which comes out to be almost 1. The mean square error (MSE) is .03 with PSNR to be 68dB.