Dott. Ing. Letizia Squarcina, Ph.D.

Post Doctoral Fellow

Fondazione IRCCS Ca' Granda Ospedale Maggiore Policlinico

University of Milan
Scientific Institute IRCCS “E. Medea”, Bosisio Parini, Italy.

NEUROSCIENCE AND PSYCHIATRY

Tecniche di analisi di MRI cerebrale
 Volume measures and diffusion in preterm infants
MRI in preterm infants:

- standard techniques do not work due to the different shape an contrast of newborn brain in respect to adult

- movement: the scans are acquired while the infant sleeps: this does not entirely prevent movement of the head,
which very heavily affects this kind of data
 Volume measures and diffusion in preterm infants
In collaboration withUOC Neuroradiologia (Prof. Triulzi, Dr.ssa Cinnante) e UOC Neonatologia
and Terapia Intensiva Neonatale (Prof. Mosca, Dr.ssa Fumagalli)

Automatic segmentation

Computation of volumes of 50 brain areas

Volume measures and diffusion in preterm infants

Cerebral total volume vs weight

Volumes at birth
vs number of injections
p = 0,684 Volume vs weight at birth
p = 0,319

Right hippocampus
Left hippocampus

450000

V
p = 0,6 p = 0,066

Right Amygdala
Left Amygdala

750 1000 1250 1500 1750

Peso_nascita (grammi)
Volume measures and diffusion in preterm infants

Number of injections
during hospitalization
90
90.00

80
80.00

70
70.00
High invasivity
60
60.00

50
50.00

40
40.00

30
30.00
Low invasivity
20
20.00

10
10.00

0
0.00
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
Volume measures and diffusion in preterm infants

Left amygdala Right amygdala

P=0.03 P<0.01

Low Punture_cutaneeHigh Low High

Punture_cutanee

Invasivity invasivity Invasivity invasivity

 Diffusion in preterm infants
In collaboration withUOC Neuroradiologia (Prof. Triulzi, Dr.ssa Cinnante) e UOC Neonatologia
and Terapia Intensiva Neonatale (Prof. Mosca, Dr.ssa Fumagalli)

DWI

T2 T1

ROI template (ALBERT) ADC

 Diffusion in preterm infants
Example:
Excluded subject due to excessive motion
Dataset:

- 30 segmented subjects

- 20 DWIs

In collaboration withUOC Neuroradiologia (Prof. Triulzi, Dr.ssa Cinnante) e UOC Neonatologia

and Terapia Intensiva Neonatale (Prof. Mosca, Dr.ssa Fumagalli)
 Diffusion in preterm infants

ADC in 50 ROI, of which the following were subjected to manual segmentation correction:

- Thalamus
Mean ADC
- Amygdala

- Hippocampus

Hippocampus Hippocampus Amygdala Amygdala Thalamus Thalamus

left right left right left right

In collaboration withUOC Neuroradiologia (Prof. Triulzi, Dr.ssa Cinnante) e UOC Neonatologia

and Terapia Intensiva Neonatale (Prof. Mosca, Dr.ssa Fumagalli)
 Diffusion in preterm infants
Preliminary results: high vs low invasivity

Mean ADC

Hippocampus Hippocampus Amygdala Amygdala Thalamus Thalamus

left right left right left right

In collaboration withUOC Neuroradiologia (Prof. Triulzi, Dr.ssa Cinnante) e UOC Neonatologia

and Terapia Intensiva Neonatale (Prof. Mosca, Dr.ssa Fumagalli)
 Machine learning
Machine learning (ML) is turning data into information

In machine learning, we look at instances of data to get knowledge

Each instance of data is composed of a number of features
The set of all the features of each instance is the feature vector
A feature vector locates an instance in the feature space

Supervised learning (pattern recognition): the system has to learn a mapping between inputs
and outputs in order to predict new values

Unsupervised learning: the system has to extract regularities in the distribution of the input
data

Classification is used to place an unknown piece of data into a known group

 Classification

Weight [g] Wingspan Webbed Back color Species

[cm] feet
1 1000.1 125.0 No Brown Buteo
Jamaicensis

2 3000.7 200.0 No Gray Sagittarius

serpentarius

3 3300.0 220.3 No Gray Sagittarius

serpentarius

4 4100.0 136.0 Yes Black Gavia

immer
5 3.0 11.0 No Green Calothorax
lucifer

6 570.0 75.0 No Black Campephilus

principalis

FEATURES TARGET VARIABLES (CLASSES)

 Classification

How to decide if a new bird is a Gavia immer or something else?  classification

The algorithm has to be trained

Training set: the set of training examples used to train the machine learning algorithms

In the bird example, the training set has six training examples. Each training example has
four features and one target variable

In a training set the target variable is known

The machine learns by finding some relationship between the features and the target
variable

In the classification problem the target variables are called classes

The number of classes is assumed to be finite
 Classification

Which are important features?

Measure just about anything that is possible to

measure and sort out the important parts later FEATURES CLASSES

INSTANCES
 Classification

Test set: dataset separated from the training set, used to test ML algorithms
In a test set the target variable is NOT known by the ML algorithm

Measure of how accurate the algorithm is  the class that the training example belongs to
is compared to the predicted value

By creating a computer program to recognize birds, an ornithologist has been replaced with a
computer
The ornithologist is a bird expert, so an expert system has been created

For a review see Veronese et al. Computational and Mathematical Methods in Medicine (2013).
 Classification

Steps in developing a classifier

Collect data

Prepare the input data

Features have to be organized in predefined formats

Analyze the input data

Is there any clearly recognizable pattern? Plotting data in one, two, or three dimensions might
help.
Not if more than three features have been measured (4D, 5D, … N-D ???)

Train the algorithm

This is where the machine learning takes place
In the case of unsupervised learning, there’s no training step (there is no target value)

Test the algorithm

 Classification

Support vector machines (SVM)

Machine learning is about constructing algorithms that can learn from data and
make decisions.

Aim is to classify, i.e. to distinguish between categories.

The algorithm learns from a training set and then applies the learnt rules to classify
elements belonging to a test set.
 Classification
Class 1

Class 2

Machine learning is about constructing algorithms that can learn from data and
make decisions.

Aim is to classify, i.e. to distinguish between categories.

The algorithm learns from a training set and then applies the learnt rules to
classify elements belonging to a test set.
 Classification

Support vector machines (SVM) represent data from the training set as points in a
multidimensional space, and finds the biggest gap that divides the two classes.

The samples that define the margins of the division between the classes are called
support vectors.

The distance between the

classes is maximized

Support vectors
Neuroimage 2015 Dec 12.. [Epub ahead of print]
 Classification: first episode psychosis patients

To distinguish healthy controls from patients with first episode of psychosis based on gray matter thickness.

We took into account possible nuisance effects of age and acquisition MRI sequences across datasets, not
correcting the data as a pre-processing step, but including the effect of nuisance covariates in the classification
phase.

To this aim, we developed a method (co-MKL) which, based on multiple kernel learning (MKL), exploits the
effect of these confounding variables with a subject-depending kernel weighting procedure

1.5T: 65 FEP (mean age +/− S.D. = 30.6 +/− 9.3 years old, 36 males) and 65 HC (mean age+/S.D. = 32.4+/−6.6
years old, 35 males)

3.0T: 62 FEP patients 29.3+/−8.9 years old, 35 males) and 62 HC (mean age +/−S.D.=29.3+/−4.3 years old, 29
males)
 Freesurfer
3D reconstruction

cortical thickness
1.5 T
Subject

3.0 T

Subject

HC
FEP

?
Confounding effects?
 Covariates!

Age MRI (1.5T or 3.0T)

HC
FEP
 Classification: first episode psychosis patients

A standard SVM classifier was applied to all the ROIs. To reduce the number of features to be analyzed with
MKL and covariate multiple kernel learning (CO-MKL), we selected the 5 best performing features with SVM.
The SVM was trained using cross-validation with a leave-one-out procedure.

We classified data acquired at 1.5T and 3.0T first separately and then together. Data were corrected for
differences in age using a general linear model (GLM).
 Classification: first episode psychosis patients
The 5 best performing features, considering both data acquired at 1.5 T and at 3.0 T, were used in a MKL
classifier, first taking the first two features and then adding the other 5 kernels one by one. Also in
this case, the nuisance effects of age as covariate were corrected using a GLM model.

Classification using simple MKL of data corrected with GLM for age differences did not improve performances
of the classifiers, reaching an accuracy of 71% for data acquired at 1.5 T and 77% for data acquired at
3.0 T.
 Classification: first episode psychosis patients

The same procedure was followed with CO-MKL. In this case, data were not corrected for the effects of age,
given that the covariates are taken into account during the training phase.

Classification using CO-MKL considerably improved the classification performances.

 Classification: first episode psychosis patients

We showed how classification performances improve when the potentially confounding influence of
covariates are taken into account instead of ignored or included in simple models as the general linear model.

In this way, we overcome the necessity to model the effect of these variables on the measured data,
so to avoid making a-priori hypotheses as the linear relationship between covariate and measure which is
assumed when using general linear models (GLM).

The classification process identified as most meaningful ROIs regions in the frontal and temporal lobes.
Prefrontal and orbitofrontal cortex represent relevant regions for psychosis and related cognitive impairments.
These areas are widely known to be involved in schizophrenia and bipolar disorder.

Taken together, these findings suggest structural, functional and metabolic prefronto-temporal changes in FEP,
which could successively lead to long-term neural and cognitive deterioration.
Dynamic Susceptibility Contrast imaging (DSC-MRI)
T2*-weighted Acquistion of a T2*-weighted sequence + tracer injection (Gadolinium)

600 Tissutal signal T2* intensity decreases at Gadolinium

[a.u.]
passage, depending on local concentration
300

0 sec.
0 25 50 75 100
Tracer concentration is given by:
70 Tissue Concentration
[a.u.]

35

0 sec.
0 25 50 75 100
Dynamic Susceptibility Contrast imaging (DSC-MRI)

After quantification of the Arterial Input Function AIF(t), estimated from arterial voxels in the
image, the following quantities are obtained which reflect the cerebral hemodynamic at a capillary
level:

1- Cerebral Blood flow 2- Cerebral Blood Volume 3- Mean Transit Time

CBF – the amount CBV – the fraction of MTT - the average

of blood that is voxel volume time that a particle
given to the brain in occupied by blood of tracer takes to
a given time pass the capillary
bed within the
voxel
Dynamic Susceptibility Contrast imaging (DSC-MRI)
In summary,
CBV

Tissutal signal Tissutal concentration

505 S(t) [au] 0.6
CVOI(t) [au]

495 0.4

DSC-MRI 485 0.2

MTT
sec sec
475 0.0
0 20 40 60 80 0 20 40 60 80

Arterial signal Arterial concentration

505 3.5
480 2.8
455 2.1
S(t) [au] AIF(t) [au]
430 1.4
405
CBF
0.7
sec sec
380
0.0
0 20 40 60 80
0 20 40 60 80
Dynamic Susceptibility Contrast imaging (DSC-MRI)

There is evidence that schizophrenia presents abnormalities in brain structure [1].

Such abnormalities, affecting structural anatomy, could be associated with anomalies in

the blood supplied to the brain.

Micro-vascular impairments may play a role in altered neuronal metabolism and

organization.

Hypoperfusion has been detected in schizophrenia using PET and SPECT [2,3].

[1] Agarwal, N., Port, J.D., Bazzocchi, M., Renshaw, P.F., 2010. Update on the use of MR for assessment and diagnosis of psychiatric diseases. Radiology 255, 23–41.

[2] Théberge, J., 2008. Perfusion magnetic resonance imaging in psychiatry. Top Magnetic Resonance Imaging 19, 111–130.

[3] Wake, R, Miyaoka, T, Kawakami, K, Tsuchie, K, Inagaki, T, Horiguchi, J, Yamamoto, Y, Hayashi, T, Kitagaki, H (2010). Characteristic brain hypoperfusion by 99mTc-ECD
single photon emission computed tomography (SPECT) in patients with the first-episode schizophrenia. Eur. Psychiatry, 25, 6:361-5.
 DSC classification in FEP

PROCESSING:

White matter / gray matter segmentation – FSL FIRST

Computation of CBF, CBV, MTT via block-circulant Singular Value

Decomposition (cSVD)

Classification using SVM and leave one out cross validation procedure

Squarcina et al. Schizophrenia Res, In Press

 DSC classification in FEP

White matter / gray matter segmentation:

 DSC classification in FEP

White matter / gray matter segmentation:

FSL
 DSC classification in FEP

White matter / gray matter segmentation:

Tissue-specific time courses

FSL

Based on the distinct gray levels of the acquired images in white and gray matter, it is
possible to segment brain tissue into white and gray matter.

This allows the distinction of blood perfusion characteristics in the two tissues.
 DSC classification in FEP

White matter / gray matter segmentation: Distribution of perfusion

indexes values

FSL

Based on the distinct gray levels of the acquired images in white and gray matter, it is
possible to segment brain tissue into white and gray matter.

This allows the distinction of blood perfusion characteristics in the two tissues.
 DSC, SVM and first episode psychosis (FEP)

We considered the following ROIs obtained from the MNI structural atlas ([1],[2]):

left and right frontal lobes

left and right parietal lobes

left and right temporal lobes

left and right occipital lobes

left and right insula

left and right caudate

left and right cerebellum

[1] Collins et al. Automatic 3-D model-based neuroanatomical segmentation. Human Brain Mapping 3(3): 190-208. (1995)
[2] Mazziotta et al. A probabilistic atlas and reference system for the human brain: International Consortium for Brain Mappi ng (ICBM). Phil. Trans. Royal Soc. B Biol. Sci. 356(1412):1293-
1322 (2001)
DSC, SVM and first episode psychosis (FEP)

The distribution of values of CBF, CBV and MTT in the ROIs were represented
with an histogram calculated over 100 discrete bins and used as feature vectors
to classify subjects into the two distinct groups of HC and FEP, after correction
for age differences
DSC, SVM and first episode psychosis (FEP)

The distribution of values of CBF, CBV and MTT in the ROIs were represented
with an histogram calculated over 100 discrete bins and used as feature vectors
to classify subjects into the two distinct groups of HC and FEP, after correction
for age differences

FSL

Squarcina et al. Schizophrenia Res, In Press

DSC, SVM and first episode psychosis (FEP)

The distribution of values of CBF, CBV and MTT in the ROIs were represented with an
histogram calculated over 100 discrete bins and used as feature vectors to classify subjects
into the two distinct groups of HC and FEP, after correction for age differences

FSL

The considered ROIs were left and right frontal, parietal, temporal and occipital lobes,
insula, caudate and cerebellum.
 DSC, SVM and first episode psychosis (FEP)

When there are more than just one feature, classical SVM does not allow the consideration of
the different features separately

Recently a different technique has been introduced, Multiple Kernel [1, 2] which allows the
simultaneous consideration of all features at the same time

Linear MK learning was applied to CBF, CBV and MTT of all considered ROIs

GLMK-SVM, considering all ROIs together, reached an accuracy of 87%

[1] Bach, F.R., Lanckriet, G.R.G., Jordan, M.I., 2004. Multiple kernel learning, conic duality, and the smo algorithm, in: Proceedings of International conference on Machine learning, ICML, pp. 41–48.
[2] Lanckriet, G.R.G., Cristianini, N., Bartlett, P., Ghaoui, L.E., Jordan, M.I., 2004. Learning the kernel matrix with semidefinite programming. Journal of Machine Learning Research 5, 27–72.
DSC, SVM and first episode psychosis (FEP)

To better understand the modification of perfusion between healthy subjects and

FEP patients, we employed a linear SVM to analyze the weights assigned to the
features.

We obtained that the most discriminant are the peak bins of the considered
histograms, i.e. there is a change in the distribution of the most frequent values
of CBV. Least discriminant are high values of CBV.

Frequency

CBV
DSC, SVM and first episode psychosis (FEP)

Application of MKL to perfusion parameters in white matter ROIs, considering all

of them together, reached 87% accuracy.

The ROIs that were given highest weight for the classification were right and left
frontal lobes, right cerebellum and right parietal lobe

Most discriminant ROIs:

- Left frontal lobe (red)

mean CBV decrease: 11%

- Right frontal lobe (yellow)

mean CBV decrease: 13%

- Right cerebellum (blue)

mean CBV decrease: 9%

- Right parietal lobe (green)

mean CBV decrease: 3%
DSC, SVM and first episode psychosis (FEP)
Application of SVM to perfusion parameters in white matter ROIs reached 83% accuracy:

Feature Accuracy (%)

Right frontal lobe CBV 83

Left parietal lobe CBV 80

Right occipital lobe CBV 79

Left temporal lobe CBV 79

Right temporal lobe CBV 77

Right occipital lobe MTT 76

Left frontal lobe CBV 76

Right caudate MTT 74

Right parietal lobe CBV 74

Left frontal lobe MTT 74

Right caudate CBF 73

Squarcina et al. Schizophrenia Res, 2015
 DSC classification in FEP

FEP can be distinguished from healthy patients on the basis of blood circulation

The most affected areas are in the frontal, temporal and parietal lobes

Considering all the regions together, FEP can be distinguished automatically from controls with a very high
accuracy, reaching 85% of cases

This confirms an involvement of vascular impairments in psychosis