In this work, we tackle the problem of early detection of Alzheimer’s disease (AD). By 2050, the number of people aged over 60 years will be raised by 1. 25 billion, constituting 22% of the global population, with 79% living in the world’s less developed. The benefits of AD early detection include -but not limited to- postponing sending the elderly to an institution, asking the elderly to take important decisions now not later, treating depression, improving caregiver’s mood, delaying the onset of dementia, and taking interventive actions. From a clinical aspect, AD can be divided into three episodes. During the first one, individuals are cognitively normal but some of them may have AD changes in brain pathology. Next comes the mild cognitive impairment (MCI), where the earliest cognitive symptoms start to occur but they do not meet the criteria for dementia. Final one is dementia, deﬁned as impairments in multiple domains that are severe enough to produce loss of functioning in the activities of daily living. The aim of this study is to apply CNN on neuroimages that are quite diverse and variable from one subject to another.
Commonly-used image datasets do not involve that level of differences between subjects belonging to the same class. In addition, we assume that our network would be versatile and competent when compared with other machine learning techniques. Methods We used the data provided by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database (adni. loni. usc. edu). Our dataset comprised 179 normal controls (NC), 134 MCI, and 85 AD. We normalized the MRI scans to the MNI reference space using the DARTEL algorithm implemented in SPM8. We selected negative amyloid-Beta protein (Aβ) NC and positive Aβ for other groups to compare against, thus the dataset became 126 NC, 96 MCI, and 75 AD. Further, we removed the effect of covariates; i. e. age, gender, and total intracranial volume (TIV) from images using linear regression. We identified all coronal slices incorporating the hippocampus, and then fed every other slice into custom CNN.
In our CNN design, all twelve scan slices defined individual channels of the input layer. The input layer was connected to two convolutional blocks, and then one fully connected layer, softmax and classification layers. Each individual convolution block consisted of convolution, maxpooling, rectified linear unit (Relu), and dropout layers. We adopted a six-fold cross validation scheme. We put the test portion apart and the rest was split into training and validation using internal four-fold cross-validation topology. More specifically, during each internal fold, after every twenty iterations, we fed forward the learned weights on the validation set; it keeps training till either validation set loss would become smaller than the previously smallest loss during five epochs or till the maximum number of fifty epochs is reached, whichever is earlier, we reshuffle the training and validation sets, and again forward feeding was repeated fifteen times; when the repetitions were complete, we started a blank network on the next fold of the six-folds.
We implemented our network using Matlab 2018a running under CentOS Linux with eight Intel Xeon CPU cores @2. 4GHz. Results What is the accuracy reached? How many iterations did the network need to converge? Which patterns did the network learn to detect AD? Interestingly, we found out that shuffling the training and validation sets and refeeding them again to the network -with previous weights as initialization- turned out to be fruitful when implemented on the test dataset in the sense that accuracy has increased. It is worth noting that increasing the number of layers beyond two resulted in deterioration in classification performance.
Using previously-trained CNN; namely, alexnet, vgg16, and resnet, resulted in poor accuracy. This is attributed to the fact that the training was done on completely different dataset of much lower variability than brain scans. Conclusions Does the approach appear useful for automated detection of AD?What remains quite challenging is the detection of subjects with subjective memory complaints (SMC); their cognitive tests are almost identical to that of normal subjects. Furthermore, optimizing the designation problem of the CNN itself; the number of layers and their sizes is quite worth to explore.
This essay has been submitted by a student. This is not an example of the work written by our professional essay writers. You can order our professional work here.