Computer assisted diagnosis of Alzheimer’s disease using statistical likelihood-ratio test

PLOS ONE RESEARCH ARTICLE Computer assisted diagnosis of Alzheimer’s disease using statistical likelihood-ratio test Xiaoming Zheng ID1*, Justin Cawood1, Chris Hayre1, Shaoyu Wang ID2, for the Alzheimer’s Disease Neuroimaging Initiative Group¶ 1 Medical Radiation Science, School of Dentistry and Medical Sciences, Charles Sturt University, Wagga Wagga, NSW, Australia, 2 Biomedical Sciences, School of Dentistry and Medical Sciences, Charles Sturt University, Wagga Wagga, NSW, Australia ¶ A complete listing of ADNI investigators can be found in the Acknowledgments * a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Zheng X, Cawood J, Hayre C, Wang S, for the Alzheimer’s Disease Neuroimaging Initiative Group (2023) Computer assisted diagnosis of Alzheimer’s disease using statistical likelihood-ratio test. PLoS ONE 18(2): e0279574. https://doi.org/ 10.1371/journal.pone.0279574 Editor: Stavros I. Dimitriadis, Cardiff University, UNITED KINGDOM Abstract The purpose of this work is to present a computer assisted diagnostic tool for radiologists in their diagnosis of Alzheimer’s disease. A statistical likelihood-ratio procedure from signal detection theory was implemented in the detection of Alzheimer’s disease. The probability density functions of the likelihood ratio were constructed by using medial temporal lobe (MTL) volumes of patients with Alzheimer’s disease (AD) and normal controls (NC). The volumes of MTL as well as other anatomical regions of the brains were calculated by the FreeSurfer software using T1 weighted MRI images. The MRI images of AD and NC were downloaded from the database of Alzheimer’s disease neuroimaging initiative (ADNI). A separate dataset of minimal interval resonance imaging in Alzheimer’s disease (MIRIAD) was used for diagnostic testing. A sensitivity of 89.1% and specificity of 87.0% were achieved for the MIRIAD dataset which are better than the 85% sensitivity and specificity achieved by the best radiologists without input of other patient information. Received: April 2, 2022 Accepted: December 11, 2022 Introduction Published: February 17, 2023 Copyright: © 2023 Zheng et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: All relevant data are within the paper and its Supporting Information files. Funding: This study was funded in part by Charles Sturt University (APC support). No additional external funding was received for this study Competing interests: No authors have competing interests. Alzheimer’s disease is the most common cause of dementia affecting ageing population in the world [1]. MRI T1 weighted structural images are recommended [2] and integrated [3] in routine diagnosis of Alzheimer’s disease (AD). The medial temporal lobe atrophy (MTA) is the hall mark of AD in MRI images [4,5]. Radiologists use a coronal section of the T1 weighted MRI images to rate patients’ MTA using a 5 points visual grading scale, based on the height of the hippocampal formation, the widths of the choroid fissure and the temporal horn [4–6]. A score of 3 or above is considered abnormal [4–6]. A diagnostic sensitivity and specificity of 85% can be achieved by the best radiologists using visual scale grading [6] without using the patients’ other information. The hippocampus is the mostly affected region among the sub-regions of the medial temporal lobe [4]. Hippocampal atrophy is one of the core biomarkers in the revised National Institute on Aging-Alzheimer’s Association (NIA-AA) diagnostic criteria for AD [7]. In addition to the atrophy of the hippocampal volume, the asymmetry of the left and right hemispheres as well as shapes and forms of the hippocampus are also of great importance [8–10], because PLOS ONE | https://doi.org/10.1371/journal.pone.0279574 February 17, 2023 1 / 11 PLOS ONE Computer assisted diagnosis of Alzheimer’s disease differences of these characteristics between AD and normal controls (NC) may point to the origin and staging of the Alzheimer’s disease [11,12]. Such differences may suggest potential targets for therapeutic drugs or directions for any drug development [13,14]. Subfields of the hippocampus have also been investigated in recent years [15]. Many machine learning (ML) algorithms have been developed and studied to assist radiologists’ diagnosis of Alzheimer’s disease using structural MRI images [16,17] including advanced deep machine learning algorithms [18,19]. The diagnostic accuracies of these algorithms are similar to that achieved by radiologists using visual scale rating, i.e. between 80–90% [16,17], or, up to 98.8% if additional information is included [18,19] for distinguishing AD from NC. These computers assisted algorithms are yet to be implemented in clinical practice, particularly, in non-research healthcare settings such as those clinics in regional and remote areas. Clinical radiologists/physicians or healthcare professionals would like to have simple yet accurate tools to assist them in their day-to-day patient care and management. The purpose of this work is to develop a simple computer assisted diagnostic tool in detecting Alzheimer’s disease using an MRI T1 weighted image without the need for knowing any other information of a patient. This is intended to be a convenient and effective tool for assisting radiologists and physicians as well as healthcare professionals such as radiographers or nurses who are caring aging populations in regional and remote areas. Methods and materials Radiologists make diagnostic decisions based on their probability knowledge of normal vs diseased images which were developed during their specialist training [20]. The decision threshold of a human observer can be biased in their decision making and an ideal observer makes decisions by placing a criterion on the axis of an underlying random variable [21]. The statistical likelihood-ratio test of an ideal observer uses the probability density functions of normal vs diseased images in its decision making [21]: LðNÞ ¼ fn ðxÞ ð1Þ LðDÞ ¼ fd ðxÞ ð2Þ and Where fn(x) and fd(x) are the probability density functions of normal (N) and diseased (D). For a decision variable x, the likelihood ratio observer uses the probability ratio of the normal and diseased or log-likelihood ratio to make decisions: y ¼ logRðD : N Þ ¼ log f ðxÞd ¼ logfd ðxÞ f ðxÞn logfn ðxÞ ð3Þ Where y>0 is considered to be diseased and y<0 is considered to be normal. The probability density functions of normal and the diseased can be constructed by using clinical data of pathologically confirmed diseased and normal patients. If the patients are randomly drawn from a population, the probability density functions fn(x) and fd(x) can be considered as normal or Gaussian distributions: � 2� 1 1 ½x mn � f n ðx (...truncated)