Advanced Machine Learning for Neuroimaging Analysis

Enabling Translational Research in NeuroImaging with Machine Learning Group Number: 58 Authors: Shanti Modi, Gayathri Ravichandran, Sushmitha Rao Uppugunduri

Background Human brain research is the most intriguing and complicated study Age and other factors have a direct correlation to its structure and behaviour Neuroimaging specialists look for measurable markers of behavior, health, or disorder to help identify relevant brain regions and their contribution to typical or symptomatic effects We have the functional magnetic resonance imaging (fMRI) and brain gray concentration matter readings to get these markers

Background Need a model that would facilitate attribution/mapping of behavior, health and disorders to regions of the brain -it s a kaggle challenge. Applying advanced analytic approaches and neuroinformatics tools for this model increases prediction accuracy for markers. We have connectivity pattern of the brain over time (temporal) and concentration of gray matter (spatial) data -more about this in datasets.

Literature Survey Machine learning has been used for many years now for diagnosis, classification of diseases and for understanding the regions of brain affecting a particular activity. For example, deep learning techniques like convolutional neural network and recurrent neural network were used to understand and diagnose early onset of alzheimer s disease from the neuroimaging data. In the paper cited at https://www.sciencedirect.com/science/article/abs/pii/S0149763412000139, support vector machine was used to identify imaging biomarkers of neurological and psychiatric disease. Support Vector machine was also used to classify schizophrenia patients from the structural MRI scans at https://www.sciencedirect.com/science/article/abs/pii/S1053811912003679.

Literature Survey fMRI or functional MRI data is used to understand the brain activity and the correlation of regions in the brain responsible for a particular activity. In https://repository.upenn.edu/cgi/viewcontent.cgi?article=1011&context=neuroethics_pubs, classifying the spatial patterns of brain activity was used for lie detection using non linear support vector machine with a Gaussian kernel. There have also been applications of CNN for understanding the brain images to predict Alzhiemer s https://www.frontiersin.org/articles/10.3389/fnins.2018.00777/full Also, LSTM based deep learning models were used to analyze the brain activity like in the case of the paper cited at https://www.frontiersin.org/articles/10.3389/fnins.2019.01321/full As we can see, lot of machine learning and deep learning models were used to analyze the functional and structural correlations of brain and also for classification and regression analysis of the subjects.

Why ML? No clear association between target variables and input features. Modelling the data distribution is complex. Therefore, cannot use simple predictive models that assume an underlying distribution. Image credits: TReNDS Neuroimaging - Data Analysis & MATLAB Files

Dataset There are three independent sets of data that are collected using structural MRI sMRI) and functional MRI (fMRI). Each of the features in dataset is obtained through Independent Component Analysis (ICA). FNC features (Functional Network Connectivity) SBM loadings (Source-Based Morphometry) SM maps (Component Spatial Maps) Target Variables : Age Domain1_var1 Domain1_var2 Domain2_var1 Domain2_var2 Train set size: 5877

Dataset Coronal Sagittal Axial Source Based Morphometry SBM Func Network Connectivity FNC Component Spatial Maps SM 53(m) * 52(i) * 63(j) * 53(k) 26 columns of Independent Component Analysis data (IC s) 1378 columns of features Each column is a correlation between two spatial maps - either two different maps or same map at different timestamp. Along dimension m, Each IC is a different part of brain (i,j) side/sagittal view, (j,k) coronal/ (top-down) view, (i,k) the axial (front-back) view Represents computational power

Preprocessing & Feature selection - FNC data : The features in this data set are the correlation values between spatial maps at different timestamps. The features were normalized to zero mean and standard deviation of 1 by using the MinMaxScaler. Training scores : The target variables of domain1_var1,domain1_var2, domain2_var1, domain2_var2 were having the NaN values which had to dropped. Feature Selection : We tried the Random Forest Regressor to plot the feature importance and then eliminated the features that were of less importance. SM features : Random slice selection -2D maps. - - -

Models SBM Loadings Random Forest, Gradient Boosting Regressor FNC Random Forest, LinearSVR, RBF Spatial maps 3D CNN Model Architecture ResNet18 with a fully connected layer at the end for prediction of 5 variables ; Total trainable parameters: ~15.5 million 2D CNN Model Architecture ResNet18 with a fully connected layer at the end for age prediction alone: ~11 million params

Individual Model Results: Though we are optimizing our model based on Mean Squared error, when we look at the individual predictions, quoting mean squared error is not relatable. Hence we mention Mean Absolute error (MAE) to understand better. FNC SBM SM mse mae mse mae mse mae age domain1_var1 domain1_var2 domain2_var1 domain2_var2

Final Model Results: Working on Integrating these models in the most optimized way .

Observations On plotting feature importance graphs in Regression models for SBM data, we see random data is contributing more than few features. Need to eliminate that data We see correlation of the data to the prediction variables is not related to the feature importance.

Next Steps 1. Integrating the optimal models for individual features 2. Build separate 2D CNN models for each of the 3 views in SM features 3. Implement EfficientNet -B4 CNN architecture for SM features 4. Implementing the CNN for understanding the FNC features

References 1. https://www.kaggle.com/c/trends-assessment-prediction/notebooks 2. https://www2.med.wayne.edu/diagRadiology/Anatomy_Modules/brain/brain.html 3. https://www.kaggle.com/c/trends-assessment-prediction/discussion/145818 4. https://nilearn.github.io/modules/generated/nilearn.datasets.fetch_neurovault_motor_ task.html 5. https://www.kaggle.com/gunesevitan/trends-neuroimaging-data-analysis-matlab-files 6. https://www.sciencedirect.com/science/article/abs/pii/S0149763412000139 7. https://www.sciencedirect.com/science/article/abs/pii/S1053811912003679 8. https://repository.upenn.edu/cgi/viewcontent.cgi?article=1011&context=neuroethics_ pubs 9. https://www.frontiersin.org/articles/10.3389/fnins.2018.00777/full 10. https://www.frontiersin.org/articles/10.3389/fnins.2019.01321/full

Thank You!