Introduction
The modern implementation of image analysis is accomplished through automated or programmable neural network architectures. Traditional configurations, such as Convolutional Neural Networks (CNN), sigmoid function input/output (IO), and multiple hidden layers are used to optimize accuracy-iteration readings. Thus, the long lasting vision of AI image recognition applications is to create a compact neural network with the lowest number of iterations (a full run- through the cycle) and accuracy to increase scalability and speed. In the case of neural networks, low iteration and high accuracy readings indicate efficiency and speed, as the system runs through the network less frequently. This concept of highly capable, scalable, and efficient neural network design creates opportunities for the predictable biomedical diagnosis through a rapid-acting learning process. Recent developments in medical-neural network architecture concentrate on single output response. Current CNN classifiers categorize images into general classes (e.g. Positive or Negative for Alzheimer’s). The current surface of research investigates radiological image diagnosis. Modern machine learning techniques provide simple positive/negative IO results to the user. In a clinically relevant setting, this provides insubstantial radiological data to the patient, radiologist, and practicing physician. Critical gaps in machine-doctor interaction are apparent due to the lack of an insightful User Interface (UI) which provides adjustable insight into diagnostic data. Perhaps the most urgent demand for early real-time diagnosis lies in Alzheimer’s Disease (or Senile Dementia).
The current trend of undiagnosed Alzheimer’s cases is marked at 50%, whereas a projected 75% of dementia patients may advance undiagnosed. A lack of early diagnosis creates precarious conditions where patients must wait on transpiring symptoms, non-unique memory tests, etc. Furthermore, early diagnosis via radiological imaging creates difficulties due to potentially false readings of undetectable plaque presence. This project investigates a combined machine learning diagnostics approach involving negative/positive IO diagnosis and heat map visualization of PET scan amyloid/tau deposits as an early indicator of Alzheimer’s disease. A five trial phase development plan is followed through each preprocessing technique. The initial testing phase consists of generalized positive, negative, and moderate PET Amyloid/tau images, processed through a final neural network system. Following five confirming test trials of confidence (0.1-0.9), preprocessing techniques are engaged consisting of image segmentation, sectioning, coloration, etc. Based on preliminary machine learning data, conjectures leading up to phase development include the following: Image sectioning via tile analysis indicates critical PET scan areas by increasing selectivity in the learning process. Specific preprocessing techniques such as gamma adjustment and edge detection may increase confidence scores by highlighting plaque-present regions of the PET scan.
An Overview of Alzheimer’s Disease and the Value of Timely Diagnosis
Alzheimer’s Disease (or Senile Dementia) is marked as a distorted form of brain degeneration which causes abnormal particles (neurofibrillary tangles and neuritic plaques) to accumulate in the brain and destroy viable neurons. The physiological effects of this disease uproot from areas which control cognition and memory. Due to the ambiguity of these physiological effects, the current aging population undergoes the barriers of a lack of diagnosis and the possibility of misdiagnosis via physician-directed evaluations. According to the Alzheimer’s association policy brief, only 45% of adults aged 65 or older diagnosed with Alzheimer's are aware of their diagnosis. Additionally, nearly 50% of patients and candidates continue their lives undiagnosed. Moreover, only 33% of the demographics diagnosed with Alzheimer’s are fully aware of their diagnosis due to ambiguous testing disclosure (Alzheimer’s Association 2015). An estimated 5.5 million Americans are currently afflicted with AD, a quantity which will triple in size by 2050 (Alzheimer’s Association 2017). The emotional and economic implications of caregiving has been highlighted by both stress and the burden of $90 billion each year (Small 2003). In addition to negating both the economic and emotional burdens of a delayed diagnosis, timely diagnosis creates an opportunity for early treatment and intervention. Ultimately, this includes an implementation of coordinated care plans, a better control of symptoms, patient safety, cost savings, and postponement of institutionalization (Saykin 2016).
Although current literature suggests the ambiguous causes of Alzheimer’s disease, molecular tracers frequently appear during radiology scans and assessments. Moreover, the accumulation of high concentrated plaques characterizes the emergence of Alzheimer’s in older adults with association to gerontological studies. Plaque accumulation is identified as beta-amyloid clumps which aggregates amyloid proteins found in the fatty membrane surrounding nerve cells. The proliferation of beta-amyloid clumps neurologically deteriorates on both a cellular and molecular level. Small clumps may block cell-to-cell signaling between neurons via synapses. Additionally, triggered immune responses may arise, activating immune cells which trigger inflammation and the devouring of disabled cells. Subsequently, the proliferation of Neurofibrillary Tangles arises- accumulations of a tau protein which collects inside neurons (National Institute on Aging 2017).
Amyloid and Tau Plaque and Tanglements
Beta Amyloid accumulation is the first pre-symptomatic step towards Alzheimer’s disease and Neurofibrillary tanglement accumulations. An overview of the beta-amyloid proteins reveals a variety of molecular forms that collect between neurons with unique levels of cohesiveness (thus, influencing severity of the disease). Amyloid irregularly produced by the breakdown of a larger protein called the amyloid precursor protein. The particular molecular form, beta-amyloid 42, is of special interest due to its unique toxicity (National Institute on Aging 2017). Beta Amyloid continually facilitates biogenesis and neurochemical pathways which ultimately lead to biomarked pathogenesis of early Alzheimer’s. Beta Amyloid deposits are primarily notorious in mediating neurochemical pathways in cholesterol homeostasis, striking Hippocampal regions of the brain. Ultimate neurochemical pathways in Alzheimer’s disease initiates eventual cerebrovascular lesions and increased neurotoxicity (Ghiso 2002). Furthermore, the amyloid precursor protein (APP) self-expresses and metabolizes in a rapidly and highly complex fashion. A primary pathway of beta-Amyloid proliferation exists through the nonamyloidogenic pathway: Full length APP is cleaved by both a- and y- secretases. The nonamyloidogenic pathway may produce several species of beta-Amyloid fragments through a cycle of fast acting cleavage and self-expression of the amyloid precursor protein.
The second stage of Alzheimer’s pathogenesis is triggered by the overexpression of Tau proteins. Tau proteins overarch and stabilize neuronal microtubules. Tau concentrates control over neuronal microtubule networks and signaling pathways which have been of direct interest in AD. Tau presents erratic molecular properties due to the nature as a natively unfolded protein. This can be demonstrated by a large number of structural conformations and biochemical alterations, including phosphorylation, proteolysis, glycosylation, etc (Mandelkow(s) 2012). Similar to aggregates of beta-amyloid protein blocks, Tau augmentation is toxic in cell and animal models, but can be reversed through suppressing protein expression or by enforcing inhibitors. In the case of AD organism models, the direct relationship between mutated/overexpressed Tau proteins and neurofibrillary tanglements exacerbates and climaxes the progression of mid-latent Alzheimer’s disease. Although Tau biomarker activity (in contrast to beta-Amyloid proteins) is elaborate in presymptomatic stages, they provide relative insight when combined with amyloid radioactive PET scans. The neuropathology of Alzheimer’s disease and dementia is directly influenced by neurofibrillary tanglement. Neurofibrillary tangles are tightly comprised of highly phosphorylated forms of microtubule tau proteins. A spontaneous disruption of the microtubule network causes disturbances with neurofibrillary tangles and phosphorylated tau species in their relative signaling pathways. According to Brion JP, a proportional relationship between neurofibrillary accumulations and and neuronal dysfunction is underlined during the progression of Alzheimer’s disease (JP 1998).
PET Scan Amyloid and Tau Image Diagnosis
The emerging generation of AD research has been transformed by the advent of carbon-11 (C-11) Radioactive Pittsburgh compound B (PiB)- a highly precise amyloid mapping tomography. The technology of amyloid Positron Emission Tomography (PET) springboarded the development of biomarkers which may potentially facilitate drug development efforts. Carbon-11 is defined with a short 20 minute half-life, creating an urgent demand for efficient and timely analysis of resulting scan images. However, substitutions, such as F-18 (flutemetamol) injections, seems more pragmatic (Johnson et al.). Beta-amyloid PET detection is implemented through intravenous injection with an amyloid PET radiopharmaceutical. However, due to PET image processing variance, a universal diagnostic standard is obscured. An existing clinical and physiological urgency for early PET scan diagnosis is critical due to a need for specificity and efficacy. Amyloid PET imaging is non-invasive and provides a direct measure of amyloid status in vivo through continuous monitoring of plaque and neurofibrillary accumulations. More importantly, the early detection or exclusion of AD is an increasing possibility with developing PET image processing. The prospective advantages of PET imaging may be useful in selecting patients for future clinical trials: Amyloid PET imaging can provide a variety of biomarkers and amyloid/tau indicators for researched therapeutic efficacy which is crucial in the current aging population.
IBM Bluemix Intelligence API and Neural Networks
IBM Bluemix is a self-assembled programming architecture comprised of cloud foundry services and IBM’s specialized Watson tools for deep learning analysis. The IBM Watson framework utilizes a variety of protocols for neural network analysis. In the context of an image classifying framework, Watson Visual Recognition is fully enforced. An overview of Watson Visual Recognition services reveals deep learning algorithms to analyze images for scenes, objects, faces, and other content. Supplementary responses may include classifying keywords that provide information about the content. The learning process simply processes example images personally uploaded by the user. Positive and negative classifiers are arranged to determine resulting confidence scores ranging from ~0.1-1.00 with potentially low threshold ratings (Gliozza et. al 2017).
General Image Preprocessing Techniques: Isolating Amyloid-Tau Image Frequencies
Image preprocessing is the cornerstone of a reliable and efficient AI Visual Recognition service, as it enhances unique features in the image. Moreover, preprocessing techniques maximize the scalability of data by effectively isolating and identifying classifying features (indicating moderate, negative, and positive diagnosis in Alzheimer’s disease). In the frame of radiological imaging, feature recognition is critical: Image variance, complexity, and anomalies become increasingly common due to larger data samples. However, a continuous preprocessing technique filters ambiguous images into absolute classifiers based on high-risk characteristics (the distribution of Tau and Amyloid plaque across the brain). Although an array of nearly 100 image processing methods are available, the projects considers edge-based detection, color inversion, gamma, and pixel sharpening as essential components of the processing procedure. Edge-based detection is the primary and wide-spread preprocessing technique used by visual machine learning frameworks. This image processing technique is used for locating common boundaries of objects within images and also recognizes discontinuities in brightness to detect general object borderlines. Edge detection is most familiar in the field of image segmentation and data extraction directly related to the areas of image processing. Frequently encountered algorithms include the Sobel, Canny, Prewitt, Roberts, and fuzzy logic methods. Typical examples, such as the “Canny” method are demonstrated below in Figure 1. The feature detection reveals the most noticeable borders of the coins through the edge detection argument/preprocessing method.