Impact of Skull Segmentation in MRI images for Alzheimer’s Diagnosis based on Transfer Learning Techniques

Abstract

This study aims to investigate how the process of skull stripping improves the performance of artificial intelligence models to predict early Alzheimer's disease through magnetic resonance imaging of the patient. The skull stripping process presented in this research was developed using thresholding, morphological, and U-net pre-trained transfer model on magnetic resonance imaging.  The experimental results confirmed that the use of the skull stripping process can significantly improve the efficiency of Alzheimer's detection via VGG16, Inception, DenseNet121, and ensemble model ResNet+Bilstm. Using skull stripping using deep learning techniques with thresholding and morphological operations produced a very clear improvement over the AI models used in this research as follows: accuracy, f1 score, and Area under curve (AUC) were improved respectively to 99.5%, 95%, and 99% for ResNet+BILSM,  up to 99.6%, 95.7%, and 99.14% for VGG16, enhance as 91%, 92.4%, and 95.47% for Inception, and improve DensNet121resultes as 79.5%, 82.2%, and 95.51%.