Data-intensive Knowledge Discovery from Brain Imaging of Alzheimer s Disease Patients

Transcription

1 PDCLifeS - EURO-PAR 2018 Invited Talk Data-intensive Knowledge Discovery from Brain Imaging of Alzheimer s Disease Patients Dr. Assoc. Prof., Head of Department Department of Computer Science University of Reading, UK G.DiFatta@reading.ac.uk 28 August 2018, Torino, Italy

2 Outline Motivations Collaborators Dementia and Alzheimer s Disease (AD) The Human Brain Structural Magnetic Resonance Imaging (MRI) Data-driven Knowledge Discovery Pre-processing: Feature generation Data manipulation Feature selection Classification A highly-descriptive predictive model Conclusions 2

3 Collaborators on this topic Many thanks to the colleagues and students I enjoyed working with: Prof. Doug Saddy, Director of Centre for Integrative Neuroscience and Neurodynamics, University of Reading, UK Prof. Mario Cannataro, Department of Medical and Surgical Sciences, School of Informatics and Biomedical Engineering, University Magna Graecia of Catanzaro, Italy Dr. Alessia Sarica, Istituto di Bioimmagini e Fisiologia Molecolare (IBFM), National Research Council (CNR), Catanzaro, Italy Dr. Alexander Spedding, PhD student, Department of Computer Science, University of Reading, UK Fabio Cappello, MSc student, Department of Computer Science, University of Reading, UK, now R&D Software Engineer at Sony Interactive Entertainment Europe, London. Miguel Sanchez, MSc student, Department of Computer Science, University of Reading, UK. 3

4 Dementia and Alzheimer s Disease Mental disorders are one of the five most costly conditions for medical spending. Dementia is the most expensive condition for medical and social cost and, still, it receives only a fraction of the research funds w.r.t. other health conditions. Alzheimer s Disease (AD): 60-80% of dementia cases Several national and international campaigns aimed at raising awareness and attention (e.g., Currently more than 40 million people suffer from AD million by 2050, leading to one of the most significant global social economic health crises 4

5 Alzheimer s Disease (AD) It is a neurodegenerative disease that slowly and progressively destroys brain cells. German neurologist Alois Alzheimer in 1907 first described the symptoms and identified the brain anomalies associated with it. AD affects memory, behaviour and cognitive function leading to confusion, changes of mood and disorientation. AD is most often diagnosed in people over 65 years. Early symptoms are mild cognitive difficulties (up to 8 years before), such as short term memory loss. In the late stage it causes a general deterioration in health and is terminal. Life expectancy following diagnosis is 3-9 years. The exact cause of AD is unknown. 5

6 Alzheimer s Disease (AD) Misconception: It is not a normal part of aging. Need for understanding causes to find better therapies and a cure. Importance of early and accurate diagnosis Quality of life can be significantly improved. 6

7 Alzheimer s Disease (AD) 7

8 Magnetic Resonance Imaging (MRI) MRI scanners use strong magnetic fields, radio waves, and field gradients to form images of the body. A scan generates a highly detailed view of tissues within the body. Structural MRI: a high resolution 3D image is generated. Brain MRI data are more frequently collected and in some cases made publicly available (e.g., ADNI, IXI).

9 Predictive Task Source data: structural MRI scans of human brains, demographic information and diagnosis of the subjects Alzheimer s Disease (AD) Mild Cognitive Impairment (MCI) It is not a type of dementia: it is a brain function syndrome with similar symptoms to AD (lesser extent). Recognised as a risk factor for developing AD Healthy Control (HC): no dementia syndrome Predictive task: Classify whether a subject is HC, MCI or AD based on a structural MRI Challenges: data manipulation, accuracy, curse of dimensionality, interpretability and confidence of models/predictions 9

10 Data Workflow data gathering, selection Image Segmentation Feature generation (stats on RoIs) data mining Pre-processing A single MRI scan file typically takes MB. After pre-processing the files generated for a single image take MB. We used MRI data of ~800 subjects: ~300GB. Pre-procesing of a single image takes ~8h ~9 months in total. A data parallel approach used to speed up computation ~10 days. 10

11 Data Workflow data gathering, selection Image Segmentation Feature generation (stats on RoIs) data mining Data Parallel task 3D brain images of 813 subjects Data Mining Predictive Task [Feature selection] Generate a predictive model from a training set: classification method Estimate accuracy of the model on a test set Use cross-validation to obtain a reliable estimate of the performance 11

12 Pre-processing Tools FreeSurfer is an open source software suite for processing and analyzing human brain MRI images. ( Image Registration, Subcortical Segmentation, Cortical Surface Reconstruction, Cortical Segmentation, Cortical Thickness Estimation, etc. The FMRIB Software Library (FSL) provides image analysis and statistical tools for fmri, MRI and DTI brain imaging data. ( 12

13 Feature Generation A total of 361 attributes: Morphological measures of specific regions of interests of left and right hemispheres: 138 cortical thickness measures (mean and standard deviation) 70 cortical areas 68 cortical volumes 16 hippocampal subfield volumes 66 volumes of subcortical structures age, gender, diagnosis volume data area data thickness data diffusion data demographical/clinical data 13

14 Data Manipulation FreeSurfer/FSL tools are based on UNIX shell scripts and generate a large number of files for a single image. Data manipulation may be a limiting factor for fast prototyping. Adopting an advanced analytics platform KNIME, the Konstanz Information Miner Introducing K-Surfer, a plugin for KNIME a tool for importing, managing and analysing FreeSurfer/FSL data in KNIME Introducing R-Surfer, an R package a tool for importing, managing and analysing FreeSurfer/FSL data in R K-Surfer

15 Data Science Platforms Gartner 2018 Magic Quadrant for Data Science and Machine-Learning Platforms 15

16 K-Surfer K-Surfer is a KNIME plug-in that contributes a number of nodes and meta-nodes to import FreeSurfer and FSL data into KNIME. K-Surfer simplifies the importing of multi-dimensional data for group analysis based on the volume, thickness and diffusion data of neuroimages. 16

17 R-Surfer Manipulating 'Freesurfer' generated data in R rsurfer is an R package that provides functionality to import the data generated by FreeSurfer version 5.3, the opensource software suite for the segmentation of brain MRIs. It helps to easily manipulate the data and provides brain specific normalisation commonly used when studying structural brain MRIs. 17

18 Feature Selection Feature selection can reduce dimensionality, thus mitigating computational performance issues and improving the classification accuracy. 1. Advanced Feature Selection composed by five stages (5S-Combo) Combining several methods and performing binary and multinominal classification: Correlation filter, Random Forest filter, SVM wrapper 2. Genetic Algorithms (GA) Exploiting the ability of GA to explore large search spaces efficiently 3. Statistical approach: differential two-step feature selection based on unequal variances T-test Exploiting similarity/dissimilarity of probability distributions in multi-label and multisource data 18

19 Advanced Feature Selection Binary classifiers: HCvsMCI, HCvsAD, ADvsMCI Multinominal OVO classification 19

20 Feature Selection Results 1. Advanced Feature Selection composed by five stages (5S-Combo), 2. Genetic Algorithms (GA) Both tested with Multinominal classification task SVM with radial basis filter kernel ADNI data 10-fold cross validation SVM used 133 features 20

21 CADDementia Computer-aided diagnosis methods based on structural brain MRI may improve the early diagnosis of dementia. CADDementia is a grand challenge workshop to objectively validate and compare methods for computer-aided diagnosis based on structural brain MRI scans. Advanced FS and SVM successfully applied to the CADDementia 2014 challenge Results of our team: 11 th /29 for multi-nominal accuracy 8 th //29 for AD classification accuracy 21

22 Classification Approaches Binary classification problem: HC vs AD Features: No selection: all the 361 numerical measurements of the brain RoIs Domain-specific feature selection: 16 hippocampal subfields Filter methods: correlation with the target variable Wrapper methods: 5S-Combo, GA algorithms Embedded methods: e.g. LASSO regression Classification methods I focus on predictive accuracy: 1. Binary Region Classifier (BRC) 2. Probability Weight-based Classification (PWC) 3. Bayesian Classifier (Bayes) 4. Support Vector Machine (SVM) Classification methods II predictive accuracy and descriptive power: Brain Age Estimation: Regression and Logistic Regression 22

23 LDA and Probability-based Classifiers PDF of volumes of 16 hippocampal subfields across subjects: HC vs AD X-axis: volume measure (v) Y-axis: P(AD v) and P(HC v) Thresholds determined by Linear Discriminant Analysis (LDA) 23

24 LDA and Probability-based Classifiers Number of subjects found positive in a number (1-14) of attributes 24

25 Classification Methods Features: 16 hippocampal subfields Classification methods: 1. Binary Region Classifier (BRC) 2. Probability Weight-based Classification (PWC) 3. Bayesian Classifier (Bayes) 4. Support Vector Machine (SVM) 10-fold cross validation (avg accuracy over ten 10-f xval trials)

26 Brain Age Estimation Can we achieve both descriptive and predictive goals? What could be a good model to predict a neurodegenerative disease with a good descriptive ability? AD makes the brain look older than the actual age of the subject. The subject age is estimated by (i) taking into account the specific predictive task and (ii) leveraging on multiple data sources. Method: 1. A context-aware regression model of the subject age: the apparent brain age 2. Classification model based on the difference between the apparent brain age and the actual age of the subject. 26

27 Brain Age Estimation Workflow D2FS brain age 27

28 Differential Feature Selection Differential Feature Selection based on Unequal Variances T- test (D1FS) Differential two-step Feature Selection based on Unequal Variances T-test (D2FS) IXI-HC D2FS ADNI-HC + T-test D1FS T-test - selected features ADNI-AD 28

29 Brain Age vs Age (male subjects) 29

30 Logistic Regression Male subjects Female subjects 30

31 Systematic Performance Analysis Classification algorithms SVM Proposed workflow with logistic regression: all possible variants of the workflow to identify contribution of each component Datasets ADNI ADNI and IXI Features All original features (brain region of interests measurements) All original features plus age deviation Only age deviation Hippocampal subfields Hippocampal subfields plus age deviation Performance indices Accuracy, specificity, sensitivity, NPV, precision 31

32 Accuracy Spotlight (M) ADNI IXI IXI data sources D1FS D2FS feature selection Apparent Brain Age single new feature 32

33 Classification Results The proposed method with a simple Logistic Regression based on a single highly descriptive feature (age deviation) achieves better accuracy than the baseline method (SVM on all features). Best accuracy achieved by SVM based on the proposed workflow for feature selection (D2FS) and the estimation of the apparent brain age. When domain-specific features are manually selected, the addition of the apparent brain age significantly improves the accuracy. Multi-source data are beneficial additional healthy subjects (IXI) improved the age regression model The two-step T-test feature selection is beneficial to improve the accuracy especially when the age deviation is used as single predictor and in all other cases as well. 33

34 Conclusions K-Surfer and R-Surfer: tools for automatic data manipulation helped to save time and to focus on the data mining challenges. With same pre-processing, normalisation and feature selection most classification methods are similar in terms of accuracy. Black-box approaches may achieve a good accuracy but may lack descriptive capabilities, especially in high-dimensional spaces. Hippocampal subfields are good predictors as known in the literature, but are not sufficient to achieve close-to-best accuracy. A new feature, the apparent brain age, helps at improving predictive accuracy in all cases and can also be used on its own for a highly descriptive classification model. Open issues Would more data be useful to improve the results further? What is the specificity w.r.t. other types of neurodegenerative diseases? Even more data would be needed to extend the work. 34

35 Publications A. Spedding, G. Di Fatta, F. Cappello and J.D. Saddy, Apparent Brain Age Estimation for Alzheimer s Disease Prediction, in preparation A. Spedding, G. Di Fatta and J.D. Saddy, "An LDA and Probability-based Classifier for the Diagnosis of Alzheimer's Disease from Structural MRI", Workshop on High Performance Bioinformatics and Biomedicine (HiBB), the IEEE International Conference on Bioinformatics and Biomedicine (BIBM), Nov. 9-12, 2015, Washington D.C., pp A. Spedding, G. Di Fatta and M. Cannataro, "A Genetic Algorithm for the Selection of Structural MRI Features for Classification of Mild Cognitive Impairment and Alzheimer's Disease", Workshop on Machine learning in decision making for biomedical applications, IEEE Int.l Conference on Bioinformatics and Biomedicine (BIBM), Nov. 9-12, 2015, Washington D.C., pp E.E. Bron et al., "Standardized evaluation of algorithms for computer-aided diagnosis of dementia based on structural MRI: the CADDementia challenge", NeuroImage, an Elsevier Journal of Brain Function, 1 May 2015, 111: A. Sarica, G. Di Fatta, G. Smith, M. Cannataro, D. Saddy, "Advanced Feature Selection in Multinominal Dementia Classification from Structural MRI Data", in the Proc. of the Workshop on Computer-Aided Diagnosis of Dementia based on structural MRI data (CADDementia), Conf. on Medical Image Computing and Computer Assisted Intervention (MICCAI), Sept. 18, 2014, Boston, MA, USA A. Sarica, G. Di Fatta, M. Cannataro, "K-Surfer: A KNIME extension for the management and analysis of human brain MRI FreeSurfer/FSL Data", Int.l Conf. on Brain Informatics and Health, August 2014, Warsaw, Poland, Springer LNCS V.8609, 2014, pp M. Berthold, N. Cebron, F. Dill, G. Di Fatta, T. Gabriel, F. Georg, T. Meinl, P. Ohl, C. Sieb, B. Wiswedel, "KNIME: the Konstanz Information Miner", Proc. 4th Annual Industrial Simulation Conference (ISC), Palermo, Italy, June 5-7, 2006, pp

36 Questions?