Repositorio Institucional de Documentos

I Acknowledgements To Jonas, my project supervisor, for all the time he has devoted me and the patience he has had with me. Thanks a lot. To Dimitri, my project director, for giving me the chance to work at EPFL. To Iñaki, for his assistance and the good advices given. To Juan Carlos, Vanesa, Elie y Jesús for the good moments spent in Zaragoza during the master and for being there whenever I needed them. To Fernando, Toño, Edu, Pilar,Eva,Pili, and all the other Master mates. You are great. To Domingo, for our mutual support in the difficult beginning in Lausanne,for the time we spent looking for accommodation, for bearing me and for all the good moments in Switzerland. To Elena, Mireia and Diego for supporting me during these months and for “burning” my laboratory on my birthday and give me a break in the last sprint. To Maria, Alberto, Maria J.,Juan and all my Erasmus mates for the luck we had being together. To my parents, my brothers and sisters and my brother-in-law for trusting me. And specially, it dedicated to the youngest in my family, Marcelo. II III Abstract Functional magnetic resonance imaging (fMRI) opens unprecedented possibilities to study the human brain at work in a non-invasive and non-radiating way. Using the hemodynamic response as a proxy for neuronal activation, this imaging modality is able to generate whole-brain volumes about every 1-4 seconds, depending on the desired spatial resolution (about 1-5 mm in each dimension). Conventional fMRI data analysis relies on a voxel-wise linear regression followed by statistical hypothesis testing. Another interesting measure that can be extracted from fMRI data is functional connectivity; i.e., for two given brain regions, the temporal correlation of their timecourses indicates their degree of "interaction". The advantage of functional connectivity is that it does not necessarily require a task and can thus be computed for resting-state scanning sessions (without any stimulation paradigm). Recently, the application of pattern recognition to neuro-imaging has gained significant interest in the field. In particular, this methodology can be applied to "decode" the subject's status (e.g., control versus disease/disorder) from functional connectivity measurements during resting state. The functional connectivity matrix is built from 90 regions as defined by a commonly used atlas. Here, we want to study the confound of the scanner on the classification result, which is an important topic to allow the generalization of this method for multicentric studies in the future. Our approach is to see whether confounds such as scanner, but also age, sex, influence the classification task. For this purpose, we have two sets of resting-state fMRI data from healthy subjects acquired using two identical MR scanners (Siemens 3T), and we have one set of resting-state fMRI from age- and gender-matched healthy subjects and patients. In the first part, we have implemented a script that automatically scans the fields in the DICOM header to check if any differences occurred. This script allows to easily screen datasets for subtle differences in the scanning protocol (e.g., echo time, repetition time) and to generate histogram of any other parameters that are stored in the data (e.g., age). In the second part, we implemented a classification pipeline with the task to decode the scanner from the data. Specifically, we showed that the scanner can be identified from the mean fMRI intensity images only, which clearly indicates that scanner-specific confounds do impact the image data. The spatial patterns that distinguish both scanners are reminiscent for subtle changes in field homogeneity of the head coils. Finally, we explore how the scanner confound influence functional-connectivity decoding, and we propose strategies to diminish this effect such as projecting the connectivity matrix in a subspace of components that do not allow to decode the scanner effect. IV V VI Resumen Functionalmagneticresonanceimaging (fMRI) es una reciente técnica no invasiva ni radiactiva que permite estudiar la actividad funcional del cerebro. Esta técnica consiste en la adquisición de una secuencia de volúmenes cerebrales y basándose en la respuesta hemodinámica del mismo permite mapear las regiones donde hay actividad. Una de las posibilidades que permite fMRI es analizar las conexiones funcionales entre las distintas regiones del cerebro tanto cuando el sujeto está en estado de reposo como durante la ejecución de una tarea. A partir de la correlación de los diferentes volúmenes adquiridos en una prueba fMRI se puede averiguar el grado de interacción que existe entre dos regiones cerebrales. Recientemente, la teoría de reconocimiento de patrones está ganado un notable interés en el campo de la neuroimagen y permite diferenciar entre estado funcional de sujetos enfermos y controles. Algunos atlas catalogan el cerebro en 90 regiones por hemisferio y con fMRI se está estudiando cómo el patrón de conectividad que posee un cerebro enfermo, en reposo o durante cualquier actividad para desarrollar su matriz de conectividad cerebral. El objetivo del proyecto fue estudiar si existen factores de confusión dificultan la tarea del conocer adecuadamente como es la matriz de conectividad durante el estado de reposo. Esto se fundamente en que previamente fue creada una matriz de conectividad para el estado de reposo pero los resultados conseguidos no fueron los esperados. Para ello se ha analizado si estos factores de confusión son debidos al escáner, la edad, el género del sujeto o simplemente a los métodos usados en la clasificación de las matrices de conectividad. Para buscar la existencia de estos factores de confusión se ejecutado dos procedimientos de estudio. Por un lado, se ha implementado un método que permite analizar las cabeceras de los archivos DICOM. Estas contienen información (tiempo de eco, tiempo de repetición, etc.) acerca de la configuración de los diferentes parámetros del escáner. Así, el objetivo fue determinar si los parámetros de dos escáneres diferentes estaban configurados de manera similar o si por el contrario había grandes diferencias entre ellos. El segundo procedimiento trata de analizar la información de intensidad sobre las imágenes de fMRI para determinar si existen cambios significativos entres éstas tanto espacial como temporalmente. Para ello se han implementado varios algoritmos basados en machine learning con el objetivo de chequear si realmente existían diferencias notables de intensidad entre las distintas imágenes. VII VIII Indice 1. Introduction .................................................................................................................. 1 1. 1. Functional Magnetic Resonance ........................................................................... 1 1. 2. Functional Connectivity ........................................................................................ 2 1. 3. Brain decoding ...................................................................................................... 2 1. 4. Functional connectivity decoding ......................................................................... 2 1. 5. Confounding factor ............................................................................................... 3 1. 6. Goal of the project ................................................................................................ 4 2. Effects of changes in acquisition protocol .................................................................... 5 2. 1. Databases .............................................................................................................. 5 2. 1. 1. CHUVrest database ....................................................................................... 5 2. 1. 2. CMSTrest database ....................................................................................... 5 2. 2. The DICOM standard ........................................................................................... 6 2. 3. Design of the DICOM checker ............................................................................. 6 2. 4. Analysis and Results ............................................................................................. 8 2.4. 1. Site-specific effects ........................................................................................ 9 2.4. 2. Group-specific changes ................................................................................ 10 2.4. 3. Effects on classification accuracy ................................................................ 11 2. 5. Conclusions ......................................................................................................... 12 3. Changes in images ...................................................................................................... 15 3. 1. Method ................................................................................................................ 15 3.1. 1. Databases ...................................................................................................... 16 3.1. 2. Preprocesing ................................................................................................. 17 IX 3.1. 3. Feature extraction ......................................................................................... 17 3.1. 4. Feature selection ........................................................................................... 18 3.1. 5. Classification ................................................................................................ 21 3.1. 6. Classification evaluation .............................................................................. 25 3. 2. Results ................................................................................................................. 25 3. 2. 1. Result on CMSTrest version A dataset ....................................................... 26 3.2. 2. Result on OHSU and WU dataset ................................................................ 33 3.2. 3. Connectivity matrix ...................................................................................... 36 3. 3. Conclusion .......................................................................................................... 40 4. Conclusions and futures steps .................................................................................... 41 4. 1. Conclusions ......................................................................................................... 41 4. 2. Futures steps ....................................................................................................... 41 Bibliography ................................................................................................................... 44 Confounding factor in functional connectivity decoding 6 acquired is of 450 by subject. CMSTrest database has been organizedinto two different datasets:  Version A: 20 patients from CMSTrest database distributed according to gender, acquisition scanner and division (EC, YC and MCI), seeTable 2.1. (a).  Version B: 24 subjects grouped by gender and group (Table 2.1. (b).). HUG BBL Male Female Male Female YC 1 0 1 0 EC 2 4 2 4 MCI 2 1 2 1 (a) Male Female YC 4 4 EC 4 4 MCI 4 4 (b) Table 2.1.Distribution of CMSTrest database.(a)Version A. (b)Version B. 2. 2. The DICOM standard The Digital Imaging and Communications in Medicine (DICOM) is a format the interchange of medical imaging.It is mainly used formanipulating, storing, picturing and transmitting these one on TCP/IP protocol. One DICOM file is structured in a header and the image acquired. The header iscomposed of severalstandard fields set which specify both administrative data (patients data, recording place, etc.) and data about the images. Some parameters or DICOM fields like the TR (Time of Repetition), TE (Time of Echo), number of slices, the proof of slices, etc. could be changed by the operator and induce confounding factors. 2. 3. Design of the DICOM checker For the CHUVrest dataset, the goal isto ascertain whether there is relationship between the successful classification of subjects and changes in the acquisition protocol fields. For the CMSTrest dataset, the goal is to study whether acquisition protocol parameters change between sites (Version A), and whether the acquisition protocol parameters change between different subject groups (Version B). Effects of changes in acquisition protocol 7 In order to perform the analysis, all DICOM fields have been read temporally by subject (time-series by patient) and then this will be comparedon all patients from the dataset. Finally,thefields whose information is variable will be analyzed to determine his relationship with the classification of subjects (CHUVrest) or scanning site (CMSTrest version A) or subject group (CMSTrest version B). The DICOM checker will receive one database of patients. At the beginning,the temporal DICOM information has been analyzed over each subject. The fields, whose values are different in each volume from the time-series, will be deleted. Normally, these are fields such asFilename, FileModDate (hour and date of the acquisition), etc. which do not give relevant information. The rest of fields by subject are stored into a table called FieldsSubj.This table has by row the different fields and by column the values for each temporal field.FieldSubj can have fields which take one value alone and another fields which take several values. The FieldsSubj tables, from the different subject, are compared and thus it can search the fieldswhose values are different among all patients. It is possible that in the FieldsSubj tables do not appear the same fields, for this, only the common fields will be compared. The criterion for filtering the non-interesting fields among subject is: - The fields whose values are different between all subjects, are not taken in account because it is not possible to establish one correlation with these ones. - The fields whose values are similar between all subjects, are not taken in account. Then overthe rest of fields a new table is built, CommonFields, where the interesting fields among all subjects from the database are stored. CommonFields has by row the interesting fields and by column the different values of this field between all subjects. FieldsSubjandCommonFieldsare used to build an integer matrix, called IntMat. This matrix will have as row the interesting fields found previously and as column the subjects that are in the database. The field i from the CommonFields takes n different values, thus an integer from 1 to nwill be assigned for this field. The process is repeated for all the fields present inCommonFields. Then the interesting fields are read over the FieldsSubj tables. The integer matrix will be built according to the values taken in the Confounding factor in functional connectivity decoding 8 interesting fields by subject and the integer assigned to CommonFields table for the same field. Fields Values fields 'ImplementationClassUID' '1.2.840.114089.1.0.0.3.1.8' '1.2.840.114089.1.0.0.3.2.2' 'ImplementationVersionName' 'DCF 3.1.8b' 'DCF 3.2.2c' 'InstanceCreationDate' '20081023' '20081124' '20081119' 'StudyDate' '20081023' '20081124' '20081119' (a) 1 1 2 1 2 1 2 1 1 3 3 2 2 2 1 3 (b) Table 1.1. (a) CommonFields table.(b)IntMat matrix. This matrix has a size of [number-fields-interest, numbersubject]. If itis taken the value IntMat(2,3), this is equal to 2. That value means that the field 'ImplementationVersionName' for the patients 3 is equivalent to 'DCF 3.2.2c' from the CommonFields table because it is in the column 2 and row from specific field. It is same for other fields of interest. Finally, the correlation is performed between the integer matrix and a vector of classification, called Labels vector. This vector will have a size of [1,N], being N the number of subjects from the database, and it will take the value 0 or 1 according to a concrete criterion desired. For example, it could get assigned the value 0 for those subjects scanned in BBL and 1 for scanned them in HUG. The correlation between the matrix and the vector described, indicates whether there is some important relationship between the interesting fields and a subject classification type. 2. 4. Analysis and Results CHUVrest and CMSTrest have been analyzed with the DICOM checker. So for each one of them has been assessed hisIntMatmatrix. Then, the Pearson Correlation Coefficient (r) between IntMat (X) and Labels (Y) vector is computed. 𝑟𝑟= 𝜎𝜎𝑥𝑥𝑥𝑥 𝜎𝜎𝑥𝑥· 𝜎𝜎𝑥𝑥 [2.1] whereσxy is (X,Y)’s covariance and σX y σY theirstandard deviations. Effects of changes in acquisition protocol 9 If correlation coefficients are less than -0.4 for a significant field and his p-value is lower than 0.05, it can be rejected that the field has a important significance. 2.4. 1.Site-specific effects The version A from the CMSTrest database will give us information about possible sitespecific effects, i.e., whether asignificant relationship is present in this database and the values taken by their parameters. The Pearson Correlation has been computed for the IntMat, calculated using the DICOM checker, from the extracteddataset and his Labels vector. This one takes the value 0, if the subject were scanned in HUG, and 1, if the subject were scanned in BBL. The process is repeated for both protocols. (a) (b) Figure 2.1.Correlation on acquisition site for fMRI protocol. (a) Correlation. (b) P-value. (a) (b) Figure 2.2. Correlation on acquisition site for MPRAGE protocol. (a) Correlation. (b) P-value. Confounding factor in functional connectivity decoding 10 The correlation and p-value on fMRI and MPRAGE are very similar (Figure 2.1 and Figure 2.2), i.e., it only changes the protocol. The number of significant fields is 25 for fMRI and 27 for MPRAGE. These fields are more or less similar in both protocol but neither of them indicates a notable correlation according to acquisition-site. 2.4. 2. Group-specific changes In this case, we want to study the scanner’s influence in the YC, EC or MCI groups. For this, we have 8 subjects per group classified in 4 men and 4 women. Labels vector takes the value 1 if the patient belongs to YC group, the value 2 if he is a member of EC group and 3 for MCI. (a) (b) Figure 2.3. Correlation on group-specific for fMRI protocol. (a) Correlation. (b) P-value. (a) (b) Figure 2.4.Correlation on group-specific for MPRAGE protocol. (a) Correlation. (b) P-value. Effects of changes in acquisition protocol 11 DICOM checker gives us 26 significant fields for fMRI protocol and 25 for MPRAGE. The OperatorName and PatientAge fields have an important correlation and this is rejected by their p-values (Figure 2.3 (b)). OperartorName field does not give a relevant information while the correlation obtained on PatientAge is normal because the group is constituted mainly by subjects with an age range from 55 to 85 years old. In CHUVrest version A there are only two subjects with 25 and 29 years old. By MPRAGE protocol (Figure 2.4), PatientAge has again an important correlation but, moreover, ImageOrientationPatienand Patientweightseem to be significant too.ImageOrientationPatien refers to the orientation of image plane acquired with respect to the patient so it does not induce influence on magnetic resonance imaging (MRI) test. Patientweight has a significant correlation by the same reason that PatientAge. Finally, for the group-specific information, no existence of confounding factors was found. 2.4. 3. Effects on classification accuracy The CHUVrestdatabase allowsto know if there is a relationship between the connectivity matrix and the scanner parameters.The Labels takes the value 0, in the patients wrong classified, or 1, in otherwise. A subject is wrong classified when a patient is a control and the classification says that he is a sick, or vice versa. The Pearson Correlation has been computed for each protocol (MPRAGE and EPI). (a) (b) Figure 2.5. Correlation on accuracy classification for fMRI protocol. (a) Correlation. (b) P-value. Confounding factor in functional connectivity decoding 12 (a) (b) Figure 2.6.Correlation on accuracy classification for MPRAGE protocol. (a) Correlation. (b) P-value. According to the accuracy results obtainedof accuracy classification, neither fMRI (Figure 2.5) nor MPRAGE (Figure 2.6) protocol correlation give significant information. In both protocols,NumberOfDicomFile shows a significant correlation but this field only says thenumber of temporal filestaken by subject. Some temporalvolumescan be missed when scanners converted the image acquired into DICOM format. This field is non-significant in based on connectivity matrix. 2. 5. Conclusions Firstly, we can observe that the fields of interest in different studies are similar. The private data elements are additional data defined by an implementer, to communicate information that is not contained in Standard Data Elements. Here, we comment somefields between different studies. From among them all it is worth the trouble to emphasize a priori: • 'ImplementationClassUID': Uniquely identifies the implementation which wrote this file and its content. • 'ImplementationVersionName': Identifies a version for an Implementation Class UID using up to 16 characters. • 'StudyDescription': Typically this field is used for a short description of the medical procedure such as you would see in CPT codes. In fact Study Description is a non-coded analogy for Procedure Code Sequence (0008,1032) where the code meaning element of the sequence item (0008,0104) is also 64 characters. Effects of changes in acquisition protocol 13 • ‘StationName’: User defined name identifying themachine that produced the compositeinstances. • 'ImagingFrequency': Precision frequency in MHz of the nucleus being addressed. • 'Private_0019_1012' and 'Private_0019_1013': These fields reference to series contract in the acquisition. • 'SpacingBetweenSlices': Defines the position of the following slices (frames) in the pixel element. Each subsequent (frame) slice from the pixel element is parallel to the first and positioned along a normal to the first slice. The normal (vector) is defined as a cross product of the row vector with the column vector of the image slice. • 'SliceLocation': Relative position of the image plane expressed in mm. See C.7.6.2.1.2 for further explanation. • 'SpacingBetweenSlices': Distance between slices. • 'ImageOrientationPatient': The image plane shall be specified by two values that designate the anatomical direction of the positive row axis (left to right) and the positive column axis (top to bottom). The first entry is the direction of the rows, given by the direction of the last pixel in the first row from the first pixel in that row.Image Orientation specifies the direction cosines of the first row and the first column with respect to the patient. • 'Private_0019_100b': It is referred to cell spacing in the acquisition. • 'RequestedProcedureDescription': Identifier that identifies the RequestedProcedure in the Imaging Service Request.Required if Sequence Item is present. • 'DeviceSerialNumber': Manufacturer’s serial number of the equipment that produced the compositeinstances. In view of all results we cannot say that there is some influence between scanners and different classifications. In each study is appreciated that the graphic results for both protocols are similar. This means the fields of interest and their values are usually the same regardless of protocols for one study. In the first study (CMSTrest version A)occurs the same than in the second study, it means there are not any significant values which give relevant information. We only Confounding factor in functional connectivity decoding 14 find some values with a high level correlation but this is not important for us. This means that the acquisition parameters stored in DICOM files are not significantly different between sites, and suggests that multi-centric studies are possible. In the second study (CMSTrest version B), where we seek the relationship between acquisition parameters and subject groups, we find interesting information a priori in:‘PatientAge’, 'ImageOrientationPatient', ‘Private_0029_1009’, ‘Private_0029_0019’, ‘PatientWeight’. The private fields are fields used by designers and it does not give us relevant information. In general, we did not find a significant influence. For the first study (CHUVrest dataset) we have a regular correlation’s graph around zero so I do not have any significant value indicating a negative effect from acquisition protocol changes in time to classify between controls and patients. In short, at the moment we cannot assure that scanners affect to different subject’s classification (CHUVrest) or that differences in acquisition protocol parameters are different between sites (CMSTrest version A). Changes in images 15 3.Changes in images The idea is to ascertain whether fMRI images coming from different scanning sites are combined to obtain larger databases, by attempting to build a scanner classifier. If we can find features in the images that let us predict which scanner an image was acquired from, then it is possible that combining data from different sites will be difficult because significant differences in some image statistics exist. A machine learning process is used to check if negative influenceexisted in the functional images data. The confounding factors have been searched spatially and temporally, i.e. thetime-series mean intensity indicates if a spatial effect is present while the variance of the temporal information searches the temporal effects. Furthermore, the confounding factors effect on temporal connectivity matrix classification will be studied. Many techniques have been proposed for statistically analysing fMRI data, and a variety of these are of general use. The aim of such analysis is to produce an image identifying the regions which show significant signal changes in response to the task. Each voxel is assigned a value dependent on the likelihood of the null hypothesis, that the observed signal changes can be explained purely by random variation in the data consistent with its variance, being false. 3. 1. Method In order to find a possible negative influence from the scanners, it is necessary to have a largedatabase of functional images. For each fMRI test, a sequence of brain volumes is acquired. This is known as the time-series of one subject. The spatial and temporal information of each subject is analyzed with a machine learning model. Figure 3.1. Machine learning blocks. Confounding factor in functional connectivity decoding 22 Performance of such systems is commonly evaluated using the data in the matrix. The following table shows the confusion matrix for a two class classifier. Actual Negative Positive Predicted Negative A B Positive C D Table 3.1.Confusion matrix. a is the number of correct predictions that an instance is negative, b is the number of incorrect predictions that an instance is positive, c is the number of incorrect of predictions that an instance negative, and d is the number of correct predictions that an instance is positive. The methods used to classify are Naïve Bayes, Support Vector Machine and Random Forest because they are models used for high dimensionality data, e.g., 86016 voxels are analyzed fromCMSTrest database version A or Bfor each subject. Naïve Bayes is an easy method of classification and useful if the population is small while Support Vector Machine is a model designed to avoid the overfitting. At last, Random Forest is a decision tree which tries to elude the variance of data. Classifications in statistical population are evaluated according to a binomial confidence interval[10].This is a range of possible proportions which may contain the true proportion. If an experiment is repeated a fixed number of times and each trial has two possible answer, the true proportion have to be inside of the confidence interval. The range is based on the 95% of the population should be good validate. • Naïve Bayes. A Naïve Bayes (NB) classifier is a simple probabilistic classifier based on applying Bayes' theorem (from Bayesian statistics) with strong (naive) independence assumptions [11]. NB is a model trained for a supervised learning which receives a training population and it learns the model to validate the testing population. Changes in images 23 An advantage of the naive Bayes classifier is that it requires a small amount of training data to estimate the parameters (means and variances of the variables) necessary for classification. Because independent variables are assumed, only the variances of the variables for each class need to be determined and not the entire covariance matrix. NB is also used as wrapper technique, Diagram 3.3 shows the block included in “Analysis Method” from Diagram 3.2 to explain how are selected the features more representatives. Inside of feature selection block is applied cross-validation. The voxelv from allthe training set (19 voxels) has been takenfor creating a new sub-training set (18 voxels) and sub-testing set (1 voxel). The sub-testing set is validated with Naïve Bayes. Finally, for the feature v is got a value between 0 and 100 in percentage indicating how well has been classified the different sub-testing sets in each fold.A feature with value equal to 100 indicates which it classifies very well and 0 which it classifies very bad. Diagram 3.3. Block Analysis Method for feature selection with Naïve Bayes. • Support Vector Machine. The Support Vector Machine (SVM) is a supervised learning methods used for classification and regression analysis and it is very useful on non-lineal problem.The SVM takes two different categories and training algorithm build a model that assigns a new sample into one category or the other. The learning statistical theory establishes two important characteristic which a learning algorithm must have: Confounding factor in functional connectivity decoding 24 - Ability to learn a rule which classifies rightly the highest number of observations in the training set. - Ability to classify rightly the new observations with that rule. Figure 3.2. The SVM algorithm [12]. The classification consists in draw a lineal boundary which divides both samples categories. In non-lineal SVM, the idea is to define a new function which builds a dimensional space where the categories are separated by one lineal hyperplane. The lineal hyperplane is chose to maximize the marginexistent between the samples of different categories. The w vector is a normal vector and it is perpendicular to the hyperplane. 𝑤𝑤= ∑𝛼𝛼𝑛𝑛𝜔𝜔𝑛𝑛𝑥𝑥𝑛𝑛 ,[ec. 3.3] where𝛼𝛼𝑛𝑛 are the support vector coefficients, 𝜔𝜔𝑛𝑛 indicates the categories labels and 𝑥𝑥𝑛𝑛 are the support vector. • Random Forest. Random forest (or random forests) is an ensemble classifier that consists of many decision trees and outputs the class that is the mode of the class's output by individual trees [3]. The basic idea is based on: - A large trees number is built without beingpruned. - Trees are built with a similar large dataset done through bootstrap on original sample. So two things are got: Correcting the error of prediction given by specific selection from the dataset. Second, having for each tree an independent sample to estimate the error of classification. Changes in images 25 - In each division of one node is selected a random variables set with a size specified and the selection of variables is limited to this sub-set. So there is a higher variability of trees and the dependence of the result with the precedent divisions is decrease. - Random Forest establishes a ranking based on the importance of the variables over the prediction of the output. 3.1. 6. Classification evaluation In machine learning, the most popular form of feature selection is cross-validation. It is an algorithm that adds the best feature or deletes the worst feature at each fold. Cross- Validation is a statistical method of evaluating and comparing learning algorithms which segments data into two groups: one used to learn a model and the other used to validate the model. In typical cross-validation, the training and validation sets must cross-over in successive rounds such that each data point has a chance of being validated against. The basic form of cross-validation is leave-one-out or k-fold crossvalidation. Leave-one-out cross-validation makes one population as testing set and the others samples as training set, while k-fold cross-validation applies k-times the validation dividing the dataset in k-samples. The concept of overfitting is important in machine learning. This generally occurs when a model is excessively complex, such as having too many parameters relative to the number of observations. A learning algorithm is assumed to reach a state where it will also be able to predict the correct output for other examples, thus generalizing to situations not presented during training. However, especially in cases where learning was performed too long or where training examples are rare, the learner may adjust to very specific features of the training data, that have no causal relationship to the target function. In this process of overfitting, the performance on the training examples still increases while the performance on unseen data becomes worst. 3. 2. Results Both databases explained in section 3.1. 1have been checked to study if there are significant changes in functional images. CMSTrest dataset allows studying the changes in images acquired in two scanners with the similar software version. Confounding factor in functional connectivity decoding 26 At the beginning, it could be supposed that theintensity mean intensity in two group of subject scanned in two different scanners is not similar. The coils from fMRI machines are not alike so the magnetic field will be different.Thus, each machine will draw with different mean intensity. If the effect on mean intensity is located spatially or temporally, it could be said that the effect could be generated by other elements like some of the antennas, etc. It is important to know if there is both spatial influence and temporal influence. The spatial influence is analyzed computing the mean intensity from the time-series while the temporal influence will be analyzed with the variance from the time-series. CMSTrest version Adataset was checked combining the different methods explained for feature selection and classification block. However, only the best features-classifier combinations are shown. After, some of them will be used to check the ADHDrestdataset. 3.2. 1. Result onCMSTrest version Adataset TheCMSTrest version Adataset has been subjected to different models of machine learning based on the features selectors and classifiers explained earlier. This are combined to corroborate the results of classificationand assuring his non-dependency of the model used. It has been built both filters and one wrapper model. The confidence interval forCMSTrest version Adataset says that the percentage of good decision has to be higher than 75%. If this percentage is lower, the lower bound on classification accuracy is random (below 50%) and it may not be said that any spatial or temporal effects exists. • Feature selection with Kruskal-Wallis and Naïve Bayes classifier (KW-NB). The intensityMean matrix fromCMSTrest version Adataset has been checked byKruskal-Wallis and it classified by Naïve Bayes. In order to select the best features, a cross-validation approach was used on the training set. Then, the testing set is classified using these features. The number of feature extracted from the training set depending on significant threshold used, named sigThreshold. Changes in images 27 To compute the threshold it must be taken in account that there are 86.016 voxels by subject. The typical significant threshold for the p-values is 0.05, but it is necessary to apply Bonferroni Correction. This correction is based on the idea that the statistical significance level (threshold) must be testing for no dependent or independent hypotheses on each dataset. Being α significance level rejecting the null hypothesis, the significance threshold has to be α/n. So, in our case, n is equal to 86.016 and the significance threshold would have to be 0.05/86.016 what is very severe. If we take a value of n less severe like 1000, the threshold will be equal to 5·10-5 and we can assess the voxel where there is a significant difference among subjects. Several thresholds have been taken to filter the non-significantvoxels. The features selected on the p-value vector given by Kruskal-Wallis for training set has been filtered with a threshold from 0.0005 to 1 with a size of step equal to 0.008. Thus the testing set has been classified based on different number of features from around 100 to 62122 voxels in mean. The remaining features to 68016 do not take in account because they are black voxel outside brain. Figure 3.3.Good decision curve CMSTrest version A dataset according to KW-NB method. (a).On IntensityMean matrix. (b)On TimeVariance matrix. The confidence interval is established in 75%. Thus, the spatial influence is guaranteed because the good decision curve on IntensityMean (Figure 3.3 (a)) is equal or higher than this interval for different number of features. On TimeVariance matrix (Figure 3.3 (b)), the classification is random for any quantity of features used, for this reason, temporal effects cannot be proved. Confounding factor in functional connectivity decoding 28 If a low threshold is taken like 0.0005 on IntensityMean matrix, an 85% of good decision is got obtainedCMSTrest version A dataset. BBL HUG BBL 8 1 HUG 2 9 (a) BBL HUG Labels 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 Predictions 1 0 0 0 0 0 0 0 1 0 1 1 1 1 0 1 1 1 1 1 (b) Table 3.2.IntensityMean matrix CMSTrest-version Aclassification with KW-NB for a threshold of 0.0005. (a) Confusion matrix. (b)CMSTrest version A dataset labels and testing setvalidation. Figure 3.4. CMSTrest-version AIntesityMean features number over classification.Threshold equal to 0.0005. Figure 3.4 shows the number of features used to validate the testing set. Subjects 1, 9 and 15 represented a substantially higher number of retained features. Interestingly, all these three subjects were uncorrectly classified (1, 9 and 15).This could be due to the high features number fed the classifier,the good decisionpercentage decreases(Figure 3.3 (a)). • Feature selection with Point Biserial Correlation Coefficient and Naïve Bayes classifier (PBC-NB). The IntensityMean matrix and TimeVariance matrix are checked withPBC to select the features on training set and then NB is used to classify the testing set. PBC let us select previously the number of features which are used to classify the testing set. The Changes in images 29 algorithm has been tried with several thresholds varying from 1 to 62122 with a step size of 75. (a) (b) Figure 3.5. Good decision curve CMSTrest-version A datasetaccording to PBC-NB method.(a). On IntensityMean matrix. (b) On TimeVariance matrix. PBC gives us the voxels ordered decreasingly according to their point-biserial coefficient (pbc). The temporal effects can be assured on 300 features (Figure 3.5 (b)) with the pbc higher, (D will be named for the number of features with pbc higher). In others cases the classification is random because it is lower than confidence interval. There are two voxels set which guarantee the existence of a spatial effect (Figure 3.5 (a)). If the testing set is validated by the voxels included in [0, 376] or [11030, 20430] according to their pbc, the classification is higher than theconfidence interval. Thus, a spatial influence exists. If D equal to 100 features, an 85% and 80% good decision on IntensityMean andTimeVariance matrix is obtained respectively. BBL HUG Labels 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 Predictions (IntensityMean) 1 0 0 0 0 0 0 0 1 0 1 1 1 1 0 1 1 1 1 1 Prediction (TimeVariance) 0 0 0 0 0 0 0 0 0 1 1 1 0 1 1 1 1 1 0 1 (c) Table 3.3. IntensityMeanandTimeVariance matrixCMSTrest-version A classification. (a) Confusion matrix, D=100. (b)Labels and prediction classification. Confounding factor in functional connectivity decoding 30 Image 3.1. Two sections for the stimulated voxels where is found a temporal influence. The color represented means the average pbc among the population sets. Yellow is pbc high and red low. The temporal influence from scanners is important in a reduced voxels setbut these are distributed in small voxels sets inside the brain area (Image 3.1.). • Feature selection with Naïve Bayes and Naïve Bayes classifier (NB-NB). The wrapper machine learning method is characterized because the system of features selection and classification is the same. Naïve Bayes classifier used as features selector by feature among the training set. Then the mostrepresentativesvoxels are used to validate the testing set with Naïve Bayes again. (a) (b) Figure 3.6. Good decision CMSTrest-version curve according to NB-NB method. (a)On IntensityMean matrix. (b)On TimeVariance matrix. IntensityMean and TimeVariance are checked for different thresholds (Figure 3.6). The spatial influence on intensity mean level is demonstrated if the n voxels which classify better are taken in the next ranges of features: [1, 411] and [7611, 46370]. The temporal influence cannot be guaranteed because the percentages of good decision only exceed the confidence boundary by the 43 voxels which represent better the training set. Changes in images 31 If a threshold of 75 and 85 are taken on IntensityMean, a 70% and 80% of good decision is obtained. BBL HUG BBL 6 2 HUG 4 8 (a) BBL HUG BBL 7 1 HUG 3 9 (b) BBL HUG Labels 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 Pred. thresh = 75 (IntensityMean) 1 1 0 0 0 1 0 0 1 0 1 1 1 1 0 1 1 1 1 0 Pred. thresh = 80 (IntensityMean) 1 1 0 0 0 0 0 0 1 0 1 1 1 1 0 1 1 1 1 1 (b) Table 3.4. IntensityMeanmatrix CMSTrest-version Aclassification for threshold of 75 and 80. (a), (b) Confusion matrix for the threshold of 75 and 80 respectively. (b) Labelsand prediction classification. • Feature selection with Point-Biserial correlation coefficient and Support Vector Machine classifier (PBC-SVM). The dataset has been analyzed with PBC as features selector and SVM as classifier. PBC-SVM also is used to check IntensityMean and TimeVariance matrix with several thresholds. In Figure 3.7, the spatial influence appears for a number of features low while the results for temporal information does not assure negative influence. Figure 3.7. Good decision CMSTrest-version A curve according to PBC-SVM method. (a) On IntensityMean matrix. (b) On TimeVariance matrix. Confounding factor in functional connectivity decoding 38 The ADHDrest connectivity matrix has been checked by PBC-SVM (Figure 3.13). The good decision percentage to validate the testing set is always higher than confidence interval boundary (67.7%). So, the classification is not random. This means the connections among WU and OHSU regions subjects are not similar. Figure 3.13. “good decision”-“Number features” curve on ADHDrest connectivity matrix and it analyzed with PBCSVM method. The regions with different connectivity among the ADHDrest subjects appear structured (Figure 3.14). Reddish regions structured (Figure 3.14 (a)) show a blue shade on Figure 3.14 (b), thus their p-values are lower than 0.05 and the ADHDrest connectivity significant differences are rejected. Image 3.5.Brodmann's Areas approximated in the human brain [20]. Changes in images 39 The 29 (Left insula), 39 (left parahippocampalgyrus), 73 (left putamen). regions taken the highest values (Figure 3. 15), so the connections on ADHDrest subjects brains regions change importantly according to the scanning site.The fact that the affected regions are always in the left hemisphere may indicate a possible scanner-related effect. (a) (b) (c) Figure 3.14. connectivity differences on ADHDrest dataset. (a)Importance map. (b) p-Values differences connectivity. (c) Important connectivity thresholded. Confounding factor in functional connectivity decoding 40 Figure 3. 15. (a) Regional feature importance connectivity on ADHDrest. The differences among ADHDrest connectivity allow validating the testing set.It means difference connectivity regions are variably structured according to acquisition site. 3. 3. Conclusion The changes the intensity searched through CMSTrest version A and ADHDrest dataset determine what spatial and temporal structures are existent. The average intensities between both scanners in each dataset (CMSTrest version A and ADHDrest) were different,but it was curious that significant regions was structured and located on region specific. Using an intensity invariant technique could determine if this confounding factor based on changes on intensity is existent. The gradient on the mean intensity of the time-series could help us with this one and it was demonstrated there were structured regions different among ADHDrest dataset. The temporal effectwas analyzed on time-variance of intensity. For this reason, the variance analysis gives spatial-temporal information.The temporal confounding factors are non-assured on CMSTrest dataset while on ADHDrest presented a significant structure located on Occipital and Parietal regions too. Finally the ADHDrest connectivity matrix was evaluated and significant differences were appreciated on the brain left hemisphere. So, we could finish saying scanners effects are influencing on fMRI images from CMSTrest version A and ADHDrest datasets. Conclusion and futures perspectives 41 4. Conclusions and futures steps 4. 1. Conclusions The neuroimaging is a young field that it advances quickly in the last years. The development of functional marker could diagnose difficult mental illness of recognizing. For making an optimal functional market is necessary to develop it with a huge quantity fMRI images. The fMRI tests for developing the functional market are taken from different scanners. This can induce a confounding factor which hampers to make good functional markers. A confounding factor can be induced by acquisition changes realized by operator or changes of intensity on images acquired. The DICOM parameters were analyzed but these do not gave relevant information. There were not significant differences on DICOM parameters from the CMSTrest or CHUVrest dataset. On the other hand, machine learning allows analyzing intensity changes on ADHDrest and CMST version A datasets. The result established spatial and temporal effects structured existed on intensity information. In an ideal case, the resting-state connectivity matrix should be similar or alike among subjects. The ADHDrest connectivity matrix was checked with machine learning algorithms and latter can validate the testing set correctly. So, functional characteristic dependent on their scanner have to exist and intensity changes could be the cause. A connectivity matrix is useful to develop functional markers but confounding factors do not allow realizing satisfactorily said task. Inferring how confounding factors effect on fMRI images, standardizing the acquisition parameters, keeping the equipment the working in order, etc. could prevent this negative effects to develop good functional markers. 4. 2. Futures steps The future objective is to remove the confounding factors. These can appear on images randomly and learning a generic method to remove them is a difficult task. For Confounding factor in functional connectivity decoding 42 developing a functional marker from huge quantity of images, the first step is to learn how the negative effect is and then this could be removed. Negative effects can be corrected combining a multiplicative (a) and additive (b) factor. 𝑥𝑥=𝑎𝑎𝑥𝑥+𝑝𝑝 b can remove the average intensity differences while the multiplicative factor will consider others cases. The correction will try each voxels independently, thus the nonaffected voxels by negative influences will be not altered. 43 44 Bibliography [1] fivan. http://fivanblog.blogspot.com/2010/02/comunicarse-con-el-pensamiento.html. June, 2011. [2] Daniel Y. Kimberg. “A functional connectivity primer”. June, 2011. http://www.ncrrn.org/papers/methodology_papers/connectivityprimer.pdf. June 2011. [3] mindpowercoach. http://mindpowercoach.net/the-brain-decoding-transmitting-beliefs. June, 2011. [4] Francisco Pereira et Al. “Machine learning classifiers and fMRI: A tutorial overview.” Int. J. Neuroimagen, 2008. [5] Jonas Richiardi et Al. “Decoding brain states from fMRI connectivity graphs”. Int. J. NeuroImagen, 2010. [6] The ADHD – 200 Sample. http://fcon_1000.projects.nitrc.org/indi/adhd200/index.html.June, 2011. [7] Statistical Parametric Mapping Manual. http://www.fil.ion.ucl.ac.uk/spm/doc/manual.pdf. March, 2011. [8]. Wikipedia. The free encyclopaedia. http://en.wikipedia.org/wiki/Main_Page. March, 2011. [9] Tom Michel. “Machine Learning”. McGraw Hill. [10] Brown, Lawrence D., Cai, T. Tony, Dasgupta, Anirban, 1999. Interval estimation for a binomial proportion. Stat. Sci. 16, 101-133.Edwin B. Wilson, Probable Inference, the Law of Succession, and Statistical Inference, Journal of the American Statistical Association, Vol. 22, No. 158 (Jun., 1927), pp. 209-212. [11] Naive Bayes classifier. http://en.wikipedia.org/wiki/Naive_Bayes_classifier. Wikipedia, April 2011. [12] Support Vector Machines. http://www.dtreg.com/svm.htm. May 2011. [13] Cross-validation. http://en.wikipedia.org/wiki/Cross-validation_(statistics). Wikipedia, May 2011. [14] Point-Biserial correlation coefficient. http://en.wikipedia.org/wiki/Pointbiserial_correlation_coefficient. Wikipedia, June, 2011. [15] Reporting activity or lesion location. http://imaging.mrc-cbu.cam.ac.uk/imaging/BrodmannAreas. June, 2011. [16] Statistic toolbox of Matlab. http://www.mathworks.com/help/toolbox/stats. March, 2011 [17] Non-Linear Dinamic. http://www.nonlinear.com/support/progenesis/samespots/faq/pq-values.aspx. March, 2011. 45 [18] “Monografias”. http://www.monografias.com/trabajos16/manual-estadistica/manualestadistica.shtml. April, 2011 [19] Columbia University Medical Center. http://www.fmri.org/fmri.htm. June, 2011. [20] Digital Imaging and Communication in Medicine. http://medical.nema.org/. June, 2011.