There are many active research projects accessing and applying shared ADNI data. Use the search above to find specific research focuses on the active ADNI investigations. This information is requested annually as a requirement for data access.
| Principal Investigator | |
| Principal Investigator's Name: | Abolfazl Torabi |
| Institution: | Students'Scientific Research Center (SSRC) |
| Department: | neuroscience |
| Country: | |
| Proposed Analysis: | the machine learning (ML) methods had been used vastly in solving the regression and classification problems in recent years and Python as a power full programming language had been a very useful tool for these purposes. In our study, we will use the ML methods using python with Alzheimer's disease data set to both classify the subjects and estimate the MCI to AD progress. The subjects can be classified into four different groups: 1. NC, 2. progressive MCI, 3. None-progressive MCI and 4. AD. Our first goal is to classify the subject into one of the mentioned groups. Then if the subject is in the progressive group, we will try to estimate the MCI to AD progress. To make this possible, different ML algorithms will be used. The data set which is going to be used is the ADNI which provides different variety of data sets including MRI images, PET scan images, biomarkers, cognitive tests, and the demographic data. MRI and PET scan images are simply 2D images that the first one contains the structural and the second one includes the functional data of brain. Biomarkers data set includes the ((((biomarkers)))) and the demographic data set contains sex, age,(((((demographic data)))) of the patients. For the classification problem, we will try deep learning methods including Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), and other classic machine learning methods like K Nearest Neighbors (KNN), Randomised KNN, Support Vector Machine Classifiers (SVM), Naive Bayes, Decision Tree and Random Forest also will be used. After training the machine with these algorithms ensemble learning methods will be used to gain better results (that in our study is to have lower False-positive score). We will use different methods for different data sets we get from ADNI. As mentioned before, ADNI will provide MRI images, PET scans, biomarkers, cognitive tests, and demographic data. In fact we will have two data sets. One the original data as mentioned above, and the other one Processed Data (PD) data set which been produced by processing original data. Some examples of these data are the volume of the hippocampus or the other areas of the brain, the thickness of the cortex, etc. There are many useful softwares developed for these purposes. Some of them are: Analysis of Functional NeuroImages (AFNI), FMRIB Software Library (FSL), Computerized Anatomical Reconstruction Toolkit (CARET), FreeSurfer, Statistical parametric mapping (SPM). We will the suitable ones for generating PD data set. CNN method will be used with PET and MRI images to classify the subjects. The CNN algorithm has some convolutional layers in it, and by using these layers, the algorithm can ‘see’ the image and get trained by them to classify the subjects. In parallel with that, the classic machine learning methods will be used on the Processed Data (PD). Recurrent Neural Networks are the best in classifying data in form of time series. This method also can be uesd to predict the value in the next time step. So the RNN will be performed on the time sequence like data such as MRI and biomarkers and cognitive tests to classify the subjects. After the classification, for those which are in MCI group, a trained RNN will be used for progress estimate purpose. Also, the CRNN which is the combination of CNN and RNN that will be tested on MRI images taken in different times to profit both RNN and CNN benefits. Other pseudo heuristic methods will be tested during this study. For example, the sequence of MRI images with 3-month steps, are in the time domain. Converting it to the frequency domain using Fourier Transform can tell us about the frequencies in the time series. Now the frequencies can be used as the training data set of classic machine learning algorithms, We have two major approaches to the progress estimate. The first one is using the sequence estimator methods like RNN to estimate the next statues given the previous ones. Each time we estimate the next status, the estimated status can be classified using the classifiers. Now in each step, we can measure the accuracy of classifiers and as it reaches a threshold, we can regard the status as converged to AD. The number of sequence estimators now can be regarded as the time at which an MCI will turn to AD. To make it clear, suppose that the data set of MRI images in ADNI are taken with 3-month steps. And the status of an imaginary subject will be AD after 4 sequence estimates. So, in fact, we can estimate that after 12 months (4 middle estimates * 3-month step for each) the status of the subject will turn to AD. The other approach which needs to be more clear in the meantime that we are waiting for the ADNI data set is using the encoders. The perceptron encoders that are widely used in feature reduction and word to vector purposes, can be used to turn our sequential data to a data set with universal time labels. We also may use other statistical analysis tools to analyze the data set. The methods mentioned so far was all Machine Learning based. So we need to bring other tools in order to make the study more accurate. For example one possibility is finding the kramers-moyal coefficients of the data set. These coefficients contains very important information. For example the first coefficient (which is called the drift term) can tell us the best prediction about the future value of the stochastic variable. Like other statistical prediction, our prediction will have an error. The second coefficient (which is called the diffusion) will tell us about the error. For other waiting time problems ( i.e. how mush to wait until read a certain output from a random process) it is common to use the backward kramers-moyal method. As an other example of the advanced statistical analysis, by statistical tools it is possible to find the potential of a random process. The local minimums of the potential will give us the attractors of the the random process. In our example the attractors are the different states of the AD that a patient can be in. the effectiveness of the mentioned methods on our specific study is not clear and will be clear after evaluating the methods on the data set. So this report of the method contains our best guesses that may work. |
| Additional Investigators |
