Low Field MRI Deep Learning Framework for Non-Hospitalizing Early Detection and Characterization of Alzheimer’s Disease Pathology

Main Article Content

Aarav Minocha

Abstract

Alzheimer’s Disease (AD) is the most common type of neurodegenerative disease and significantly disrupts the func- tioning of the brain, leading to memory loss, cognitive decline, and more. Currently, there is no effective method to cure or treat AD, making an early diagnosis significantly important so preventative actions can be taken before large scale degradation occurs. The most common diagnostic techniques consist of various neuroimaging modalities, which provide significant structural and functional information within the brain. Many current diagnostic techniques regarding neuroimaging are extremely expensive, have long scan times, and require the patient to have access to a hospital with high-grade equipment such as MRI, PET, and CT scanners. One emerging technique within current research to address these problems is the use of Low-Field MRI scanners, which are cheap, transportable, and have much shorter scan times compared to traditional MRI scanners due to their low Tesla. However, these scanners provide a significantly less amount of information when compared to expensive hospital-grade MRI scanners because of their low magnetic strength. Therefore, the goal of my study was to create an image enhancement framework to effectively segment the volume of regions associated with AD (Amygdala, Hippocampus, and Ventricles) and automatically diagnose AD utilizing Low-Field MRI scans. I created a framework to generate synthetic Low Field (LF) MRI scans using a Fourier Transformation framework. I then developed a deep learning framework to enhance LF MRI scans utilizing SRCNN and UNET++ deep learning models which super resolved and segmented the scans, respectively, to obtain volumetric information. My SRCNN achieved a Mean Squared Error of 214.54, a Peak Signal to Noise Ratio of 31.2 dB, and a Structural Similarity Index Measurement of 0.82. My UNET++ achieved an Accuracy of 96.3, a Precision of 89.3, a Recall of 85.6, and a Dice Score of 0.93. Finally, I utilized the volumetric information to classify scans into AD or normal utilizing a soft voting majority framework, where the final diagnosis was the one majority of models agreed with. My Soft Voting framework consisted of three machine learning models (Linear Regression, Support Vector Machine, and Multilayer Perceptron), and achieved an accuracy of 0.96, a precision of 0.96, and a Recall of 0.98.

Article Details

Section
Articles