Guidelines for Effective Automatic Multiple Sclerosis Lesion

Segmentation by Magnetic Resonance Imaging

Giuseppe Placidi

1 a

, Luigi Cinque

and Matteo Polsinelli1

VI-Lab, c/o Department of Life, Health and Environmental Sciences, University of L’Aquila, Coppito 2 AQ, 67100, Italy

Department of Computer Science, Sapienza University of Rome, Via Salaria 113 RM, 00198, Italy

Keywords:

Image Identiﬁcation, Image Segmentation, Multiple Sclerosis, MRI, Convolutional Neural Networks.

Abstract:

General constraints for automatic identiﬁcation/segmentation of multiple sclerosis (MS) lesions by Magnetic

Resonance Imaging (MRI) are discussed and guidelines for effective training of a supervised technique are pre-

sented. In particular, system generalizability to different imaging sequences and scanners from different man-

ufacturers, misalignment between images from different modalities and subjectivity in generating labelled im-

ages, are indicated as the main limitations to high accuracy automatic MS lesions identiﬁcation/segmentation.

A convolutional neural network (CNN) based method is used by applying the suggested guidelines and pre-

liminary results demonstrate the improvements. The method has been trained, validated and tested on publicly

available labelled MRI datasets. Future developments and perspectives are also presented.

1 INTRODUCTION

Multiple sclerosis (MS) is a degenerative disease of

the brain and spinal cord which can vary greatly be-

tween patients in severity and symptoms (Steinman,

1996). The majority of patients transit into a progres-

sive phase consisting in an unremitting and progres-

sive accumulation of disability. Actually there is no

cure for MS and existing therapies focus on symp-

tomatic management and prevention of further dam-

age, with variable effectiveness, though recent ad-

vancements are promising. MS origins are not well

understood but characteristic signs of tissue damages

are recognizable, such as white matter lesions and

brain atrophy or shrinkage due to degeneration. These

signs can be observed by MRI which is a special tool

to follow-up MS patients with reduced invasiveness

due to the usage of speciﬁc contrast agents. In fact,

focal lesions in the brain and spinal cord are primar-

ily visible in the white matter on structural MRI ob-

servable as hyperintensities on T2-weighted images,

proton-density images (PD), or ﬂuid-attenuated inver-

sion recovery images (FLAIR), and as hypointensi-

ties, or “black holes”, on T1-wheighted images (Fil-

ippi et al., 2019). These imaging procedures are all

performed in a single MRI examination and the corre-

sponding slices (hundreds) are all used for MS moni-

toring and follow-up (also comparisons with previous

examinations are necessary). Identiﬁcation of the le-

https://orcid.org/0000-0002-4790-4029

sions affecting the white matter and their count and

volume calculation by MRI have become well es-

tablished protocols for assessing the disease progres-

sion and pharmacological efﬁcacy. For this reason,

MRI is currently used routinely in clinical practice:

imaging markers are capable to capture volumetric

changes but need to be assisted by an expert, either

human or automatic. However, the richness of MRI

parameters/imaging modalities if, by one side, con-

stitutes an advantage for gathering fundamental infor-

mation about MS lesions, by the other it makes the de-

sign of efﬁcient automatic experts a real challenge be-

cause images and, hence, the corresponding features,

change with magnetic ﬁeld strength, imaging parame-

ters, sequences and scanners from different manufac-

turers (Siemens, Philips, GE, etc.). To these modiﬁ-

cations, a trained human eye suddenly adapts but an

automatic expert has to be deeply trained before its

adaptation. But, is this really necessary?

In what follows we describe some guidelines

for automatic segmentation of MS lesions identiﬁca-

tion/segmentation by MRI and discuss how to allow

an automatic system to perform at best. Moreover, we

present a strategy to improve lesion identiﬁcation and

segmentation. To the best of our knowledge, the pro-

posal of preliminary conditions for correct MS lesion

identiﬁcation/segmentation by MRI is new and nec-

essary to obtain better performance from automatic

methods. The manuscript is structured as follows:

Section 2 provides the related work, Section 3 dis-

cusses some critical points and presents the guide-

570

Placidi, G., Cinque, L. and Polsinelli, M.

Guidelines for Effective Automatic Multiple Sclerosis Lesion Segmentation by Magnetic Resonance Imaging.

DOI: 10.5220/0009150705700577

In Proceedings of the 9th International Conference on Pattern Recognition Applications and Methods (ICPRAM 2020), pages 570-577

ISBN: 978-989-758-397-1; ISSN: 2184-4313

lines, Section 4 shows the effects of using the pro-

posed guidelines on a benchmark automatic method

and Section 5 concludes the paper and presents future

developments.

2 RELATED WORK

MRI is considered the gold standard for identiﬁca-

tion and evaluation of MS lesions affecting white mat-

ter, thanks to its richness ofparameters, which allow

to highlight lesions with respect to the healthy tissue

also by using speciﬁc contrast agents (Di Giuseppe

et al., 2001; Placidi, 2012). Hundreds of MRI slices

composing a single 3D examination are usually anal-

ysed by expert radiologists: the operation is time

consuming, subjective and difﬁcult to be carried out

without errors due to the huge number of evaluations

and comparisons required for each of the identiﬁed

lesions. Moreover, additional evaluations and com-

parisons are required between the current examina-

tion and previous ones to quantify the progression of

disease (Placidi et al., 2003). In fact, both the num-

ber of new lesions occurring after the last control and

the increment in extension of the old lesions are use-

ful parameters to deﬁne the status of disease. This

pushes the development of automatic lesion identiﬁ-

cation/segmentation strategies.

Regarding automatic segmentation of MS lesions

by MRI, several attempts have been done with suc-

cess, though the variability of MS lesions in size,

shape, intensity and localization make automatic and

accurate identiﬁcation and segmentation really chal-

lenging (Garc

ıa-Lorenzo et al., 2013; Danelakis et al.,

2018; Commowick et al., 2018). Though classical

techniques, based on shapes (Franchi et al., 2009;

Maurizi et al., 2009), could be effective, a partic-

ular attention to deep neural networks is necessary,

due to their accuracy in solving computer-vision tasks

with low manual intervention with respect to other

approaches. The great advantage of deep learning is

that the feature set would be no longer deﬁned by the

user but learned directly by the system from the train-

ing images. This is a useful property because it is

often difﬁcult for people to characterize features that

best serve to separate healthy tissue from MS lesions.

From the perspective of deep learning application, the

high dimensionality of the MR images, the difﬁculty

of obtaining reliable ground truth and the high ac-

curacy required for clinical practice, all contribute to

make MS lesion identiﬁcation/segmentation a worthy

test application. CNN have demonstrated breaking

performance also in brain imaging segmentation (Yoo

et al., 2014; Vaidya et al., 2015; Valverde et al., 2017).

In particular, Yoo et al. were the ﬁrst to propose an

automated learning approach for MS lesion segmen-

tation. Besides the architecture of the used system,

the interesting innovations were that 3D patches of

the MRI volume were used. In 2015, Vaidya et al.

proposed a method that used 3D CNNs to learn fea-

tures by different datasets of the same patient: T1-

w, T2-w, PD and FLAIR MRIs. The method pro-

posed in (Valverde et al., 2017) has proven to use ef-

ﬁciently the information carried on by different MRI

imaging modalities by reducing the number of param-

eters (and hence the training set) through the usage of

two CNNs in cascade, trained separately. To date, the

method presented in (Valverde et al., 2017) represents

for MS lesion segmentation one of the benchmark

architectures. In fact, a comparative study of algo-

rithms for MS lesion segmentation for MICCAI2016

international challenge (http://www.miccai2016.org),

presented in (Commowick et al., 2018), demonstrated

that the method in (Valverde et al., 2017) was estab-

lished as one of the most effective for MS lesion seg-

mentation, though the best method was that obtained

by creating a consensus between the results of all the

compared methods. However, though advanced com-

puter vision techniques have been compared in (Com-

mowick et al., 2018), the results were modest with re-

spect to other ﬁeld of applications. In what follows we

discuss the reasons of poor results and suggest guide-

lines to allow better efﬁcacy for automatic strategies.

3 GENERAL CONSIDERATIONS

AND GUIDELINES

DEFINITION

Though MRI is considered a gold standard, the cor-

rect interpretation of MS lesions through MRI is still

a subject of debate (Filippi et al., 2019) due to the

fact that MS lesions can be easily misdiagnosed or

erroneously interpreted (confused with other diseases

and/or artifacts and/or tissue modiﬁcations with age)

also by expert, trained radiologists and guidelines for

radiologists are continuously updated to overcome

misdiagnosis (Filippi et al., 2019; Thompson et al.,

2018). Moreover, in (Filippi et al., 2019) it is also

afﬁrmed that misdiagnosis also depends on the used

MRI scanner. As a consequence, expert radiologists

often disagree when performing independent diagno-

sis of the same data, both due to the ambiguity be-

tween MS lesions and other diseases and because

they could have gathered their experience on different

scanners. This disagreement is conﬁrmed in (Com-

mowick et al., 2018) where data contained into the

Guidelines for Effective Automatic Multiple Sclerosis Lesion Segmentation by Magnetic Resonance Imaging

571

MICCAI2016 dataset from 53 patients were inter-

preted by 7 independent radiologists: the resulting la-

belled version of the MRI images were obtained by

producing a consensus between them. This repre-

sent the ﬁrst fact which distinguish MS interpretation

with respect to other computer vision problems: the

problem is not uniquely deﬁned. This, obviously, re-

ﬂects on the performance of any potential automatic

strategy because also the ground-truth used for train-

ing could confuse it. In addition, other important

considerations have to be done regarding data them-

selves for which we continue to refer to the MIC-

CAI2016 dataset (Commowick et al., 2018), being

one of the most important benchmark datasets actu-

ally available to test automatic MS lesion identiﬁca-

tion/segmentation strategies. MICCAI2016 dataset

is composed by MRI images collected with different

imaging modalities (PD,T1-w, T2-w and 3D FLAIR),

from different centres, with 4 different scanners per-

forming at different magnetic ﬁelds (one at 1.5 T and

three at 3T), by three different manufactures (Philips,

Siemens and GE). Data for the challenge were divided

in two groups: those from 15 patients were furnished

labelled to the participants to train their methods (at

the end of the process, the participants were asked to

provide their code for internal test); those from 38 pa-

tients were maintained secret and used by the person-

nel of the challenge to evaluate the performance of

the methods participating to the challenge. The ﬁrst

group contained patients from all the centres and from

all the scanners except one, the GE scanner, whose

data were maintained obscured to verify the robust-

ness of the algorithms when using a scanner differ-

ent from those used for training. Data were furnished

both in unpreprocessed and in preprocessed form.

The unpreprocessed form consisted of raw data, as

produced by scanners, while preprocessed data con-

sisted in performing the following steps:

• Denoising of each modality;

• Rigid registration of each modality on the FLAIR

image;

• Brain extraction (skull stripping) from T1-w im-

age and applied to other modalities;

• Bias ﬁeld correction of each modality.

For the methods used in different steps, please re-

fer to (Commowick et al., 2018). Any participant

group, for its strategy, was free to use or not the all

imaging modalities and to choice between unprepro-

cessed or preprocessed data. In fact, some of them

decided to use a reduced set of imaging modalities

and/or to use unpreprocessed data. The evaluated 13

identiﬁcation/segmentation strategies were all tested

by using F1-score and Dice-score, the ﬁrst to test the

capability in identifying a lesion and the second for

measuring the capability in segmenting correctly a le-

sion (being both identiﬁcation and segmentation nec-

essary parameters to establish the progression of the

disease). Results demonstrated that any of the pre-

sented methods performed worse than the worst hu-

man radiologist (compared with the ground-truth ob-

tained by merging the identiﬁcation/segmentation by

the 7 radiologists) both in F1-score and in Dice-score

that these performances got worse when the methods

were tested on data from the secret scanner (GE), on

which no strategy was trained before. The identiﬁ-

cation/segmentation results slightly improved if the

output of all strategies were merged in a consensus:

in this way, the results were almost comparable with

those of the worst human expert.

Besides the considerations in (Commowick et al.,

2018), some important aspects have to be under-

lined (Roy et al., 2018):

1. MS lesion identiﬁcation/segmentation depends,

among other factors, on imaging scanners due to

differences in imaging parameters, temporization,

features, magnetic ﬁeld values and homogeneity,

etc., which could have more inﬂuence on auto-

matic methods than on human experts because hu-

mans use also other implicit information (clinical

or anatomical concepts, etc.) to evaluate the im-

age content: a huge increment of data for train-

ing should be necessary to include differences be-

tween scanners into an automatic system;

2. MS lesion identiﬁcation/segmentation depends

on the used data pre-processing strategy which

should be part of the method itself: the indis-

tinct free usage of data (preprocessed or unprepro-

cessed) could greatly affect the convergence of the

method and the training dataset dimension;

3. An MS lesion identiﬁcation/segmentation strategy

depends on the imaging modalities it uses (FLAIR

and T2-w images are more informative than PD

or T1-w (Filippi et al., 2019): the indistinct usage

of all the modalities to train an automatic strategy

probably results in a decrement of convergence

speed and has to imply an increment of the dataset

used for training.

The previous considerations found their conﬁrma-

tion in the contrasting results reported in (Commow-

ick et al., 2018): the methods performance decreased

when used on data from a previously unseen scanner;

methods which used preprocessed data were not all

better than those using unpreprocessed data; methods

using all the imaging modalities were not always bet-

ter than those using just some imaging modalities.

To better explain these apparently strange be-

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

572

haviours, please consider data presented in Figures 1

and 2, where some images, from the MICCAI2016

dataset, collected by different scanners are reported

for all the imaging modalities, both before (Figure

1) and after preprocessing (Figure 2). For the same

images, an horizontal line of data (red line) is also

plotted below (Figure 1b and Figure 2b). As can

be noted, unpreprocessed data show relevant differ-

ences between scanners (though data allowed to dif-

ferent patients, it is clearly visible the ratio between

the amplitude of different tissues in the same image

are different for the two scanners, as it is also con-

ﬁrmed by comparing the image corresponding to the

same imaging sequences): these differences, which

distinguish MRI from CT (where images from differ-

ent scanners are scalable in amplitude and easily com-

pared), are due to different imaging parameters opti-

mization by different manufacturers, though using the

same imaging sequences.

In Figure 2, the situation after preprocessing, an

amplitude normalization between different images

has occurred. In fact, the images of different scan-

ners are more similar than those before preprocess-

ing. However, from Figure 2b it can be observed

that the preprocessing step produced a variation on

the baseline of some of the images (the signal out-

side the brain, which should be zero, has a level well

above zero). Moreover, each image was normalized

independently from the other: this implied a modi-

ﬁcation which has been different from one image to

the other, thus introducing substantial differences also

on data from the same scanner. Finally, the ampli-

tude ratio between different tissues in the same im-

age has not been rightly corrected and, in some cases,

differences between data coming from different scan-

ners were increased. This is probably the reason why

some automatic strategies, though using preproessed

data, performed worse than those using original, un-

preprocessed, data. Finally, from both Figure 1 and

2, it can be noted that the information carried on by

different imaging modalities regarding MS lesions is

completely different: iperintense regions on FLAIR

images which are also iperintense on the correspond-

ing T2-w images surely indicate MS lesions (Filippi

et al., 2019). The other imaging modalities (T1-w and

PD) do not add anything more and, often, their con-

tent is confusing and not clearly interpretable (as in

the MS lesions indicated by the green arrows, both in

Figure 1 and Figure 2).

Form the above considerations, the following

guidelines could be derived:

1. The training of the method should be done on data

from a single scanner (also humans adapt to the

scanner they normally use): when data from dif-

ferent scanners need to be interpreted and, may

be, compared, the system has to be trained sepa-

rately to each scanner (in this way, the training set

can be reduced, the procedure shortened and the

performance increased);

2. A preprocessing strategy, consisting in the rigid

registration of each modality on the FLAIR im-

age, is necessary to obtain images of differ-

ent modalities which are spatially correspondent.

Other forms of preprocessing, especially those

consisting in amplitude corrections, have to be

performed on the whole volume and not differ-

ently on each single slice. Moreover, preprocess-

ing has to become part of the automatic segmen-

tation method;

3. The image modalities to be used in the identiﬁ-

cation/segmentation process have to be chosen in

advance to avoid useless/confusing information,

unjustiﬁed increment of the training dataset, con-

vergence deceleration and performance reduction

(FLAIR and T2-w images are sufﬁcient).

In what follows, we show how, by applying

the previously deﬁned guidelines, it is possible to

improve the performance of a lesion segmentation

method.

4 MS LESION IDENTIFICATION/

SEGMENTATION

Being a benchmark method, we have used the super-

vised CNN-based paradigm presented in (Valverde

et al., 2017) that has also been used, in a modiﬁed

version, in (Placidi et al., 2019). In particular, by fol-

lowing the previously deﬁned guidelines, we operated

the following choices:

1. the dataset used for training, validation and test

was the MICCAI2016 dataset but just using data

from a single 3T scanner (Philips manufacturer);

2. raw, unpreprocessed, data were preprocessed by

performing rigid registration of each modality on

the FLAIR image followed by brain extraction

(skull stripping) from T1-w image and applied to

other modalities;

3. only FLAIR and T2-w imaging modalities were

used for identiﬁcation/segmentation. In this way,

we provided a simpler task to the system, thus

reducing the dimension of the training, labelled,

dataset. The images selected from the dataset

were distributed in three subsets: 800 for train-

ing, 200 for validation and 100 for test. A scheme

Guidelines for Effective Automatic Multiple Sclerosis Lesion Segmentation by Magnetic Resonance Imaging

573

FLAIR

PhilipsSiemens

PD T1 T2 FLAIR

PhilipsSiemens

Figure 1: Raw, unprocessed, data from different scanners (rows) and from different imaging modalities (columns). Images

are reported in (a) and plots of a single row of the images (along the red line) are shown in (b). The position of a lesion along

the red line is indicated by an arrow. The shrinkage of the FLAIR image from Siemens scanner is due to a different (greater)

dimension of the voxel in the horizontal direction.

of the assembly used for MS lesion identiﬁca-

tion/segmentation is reported in Figure 3.

The method is based on a cascade of two CNNs.

The low variation in contrast of MRI images, the use

of images from just one scanner and the reduction of

imaging modalities, allow simple network architec-

tures and a reduction of the training set dimension.

The system consists of a 7-layers architecture for each

of the two CNNs. Each network is composed by two

stacks of convolution and max-pooling layers with 32

and 64 ﬁlters, respectively. Convolutional layers are

followed by a fully-connected layer of size 256 and

a soft-max fully connected layer of size 2 whose out-

put is the probability of each voxel to belong to a le-

sion. For a complete settlement of the used parame-

ters, please refer to (Valverde et al., 2017). MS lesions

are calculated using 3D neighboring patch features.

The used 3D patches are cubic, 11x11x11 voxels. The

splitting in two different CNNs allows to separate the

training procedure in two and this allows a reduction

of the number of parameters without reducing accu-

racy. To reorder data balance for training, that is to

equilibrate the number of “positive” patches (contain-

ing lesions) with “negative” patches (containing no le-

sions, much greater than the other), the dataset used

for training consists of the whole dataset of positive

patches and of an equal number of randomly selected

negative, healthy patches. In this way, the ﬁrst net-

work (CNN1) is trained by using the resulting bal-

anced dataset and then tested on the whole dataset,

thus obtaining a list of probabilities for each voxel of

each patch to be “positive” (part of a lesion). Af-

ter that, a balanced dataset is created by using the

previous test results and by considering as positive

all patches containing voxels whose probability is

greater than 0.5. As for the previous balanced training

dataset, negative patches (those in which all voxels

had probability <0.5), are randomly selected to be the

same number of “positive” patches. The second net-

work (CNN2) is trained from scratch with the dataset

resulting from CNN1. Once the whole pipeline is

trained, new unseen MRI volumes can be processed

using the same, two stage, architecture. The dataset

is ﬁrst decomposed in patches and, then, all volume

patches are evaluated using CNN1. CNN1 discards

all voxels with low probability (< 0.5). The rest of

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

574

PD T1 T2 FLAIR

PhilipsSiemens

PD T1 T2 FLAIR

PhilipsSiemens

Figure 2: Data of Figure 1 after preprocessing. Images are reported in (a) and plots of a single row of the images (along the

red line) are shown in (b). The position of a lesion along the red line is indicated by an arrow. Images have been also reshaped

after their co-registration.

the voxels, included into corresponding patches, are

re-evaluated by CNN2 to obtain the ﬁnal probabilis-

tic lesion mask. Resulting binary masks (ones where

lesion are present, zeros elsewhere) are computed by

thresholding the probability lesion masks (prob > 0.5

are considered lesions). Finally, an additional false

positive reduction is performed by discarding binary

connected regions with very low number of positive

voxels (this number is calculated with respect to the

minimal volume of the lesions used for testing). The

method had an average F1 score of 0.68 and an av-

erage Dice score of 0.71 (about 25% better than the

original method (Valverde et al., 2017) and 15% bet-

ter than the modiﬁed method in (Placidi et al., 2019)

without using any artiﬁcial strategy for increasing the

training dataset of patches. The improvement with

respect to (Placidi et al., 2019), relevant if we con-

sider that it has been obtained with half of the imag-

ing modalities, is mainly due to the fact that it has

been obtained by training the method on data from

a single scanner and just from the most signiﬁcant

imaging modalities, which simpliﬁes the identiﬁca-

tion/segmentation process. Moreover, these results

are signiﬁcant because they allow to overcome the

score of the automatic ”Team fusion” and also of the

worst human expert (Commowick et al., 2018), thus

making automatic identiﬁcation/segmentation accept-

able for MS diagnosis/analysis. In order to show the

results on the images, Figure 4 reports the worst-case

automatic identiﬁcation/segmentation: the method al-

lows a discrete identiﬁcation of the lesions (false pos-

itives are in red) and a good segmentation (false neg-

atives are in blue).

5 CONCLUSION

We have discussed some limitations that occur when

using automatic identiﬁcation/segmentation of MS le-

sions by MRI data: the richness of imaging param-

eters and internal variability of MRI scanners make

the problem ambiguous and difﬁcult. By consider-

ing these limitations we have extracted a set of basic

guidelines that the training dataset should have in or-

der to avoid confusion when training a supervised au-

tomatic identiﬁcation/segmentation strategy. Finally,

we have applied these guidelines and used them while

performed training of a CNN-based strategy used as

a benchmark. The results are better than those ob-

tained without using the constraints on the training

Guidelines for Effective Automatic Multiple Sclerosis Lesion Segmentation by Magnetic Resonance Imaging

575

T2 W

CNN1

Conv – pool

layers

Fully

connected

Image

Patch

Flair

Image

Patch

CNN2

Conv – pool

layers

Fully

connected

Training

dataset

Figure 3: Two stage CNNs architecture used for identiﬁcation/segmentation of MS lesions. Input of the system are the

registered volumes by FLAIR and T2-w images. Training of CNN2 is made with a separated dataset.

a) b)

Figure 4: MS lesion identiﬁcation/segmentation on one of

the images (FLAIR) by MICCAI2016 used for test. In (a),

the ground-truth identiﬁcation/segmentation is reported in

green; in (b), the same image is reported with indicated, in

colors, the voxels identiﬁed/segmented by the method: the

voxels rightly identiﬁed/segmented are indicated in green;

in red are those wrongly identiﬁed as lesions (false posi-

tive); in blue those are those wrongly recognized as healthy

tissue (false negative).

dataset, thus making the automatic method similar, in

performance, to a human expert. Moreover, we have

obtained a faster convergence of the method with re-

spect to use it with data from multiple scanners and/or

when using data from indistinct imaging modalities.

Future work will be dedicated to train the method

also by using the other imaging modalities in order to

test the effective usefulness of these modalities in the

MS identiﬁcation/segmentation process. Moreover,

the method will be also trained on data coming from

different scanners in order to quantify the contribu-

tion of the scanner on the identiﬁcation/segmentation

process and to verify if the method is generalizable to

different scanners. Finally, due to the reduction of the

dimension of the problem (data from just one scanner

and from two imaging modalities) we want to inves-

tigate the optimization of the method by proposing a

CNN-based approach deﬁned on the whole image and

not on patches: this would be preferable for training

the method also regarding the position of the lesions

inside the image (white matter) and to reduce outliers.

REFERENCES

Commowick, O., Istace, A., Kain, M., Laurent, B., Leray,

F., Simon, M., Pop, S. C., Girard, P., Ameli, R., Ferr

J.-C., et al. (2018). Objective evaluation of multiple

sclerosis lesion segmentation using a data manage-

ment and processing infrastructure. Scientiﬁc reports,

8(1):13650.

Danelakis, A., Theoharis, T., and Verganelakis, D. A.

(2018). Survey of automated multiple sclerosis le-

sion segmentation techniques on magnetic resonance

imaging. Computerized Medical Imaging and Graph-

ics, 70:83–100.

Di Giuseppe, S., Placidi, G., and Sotgiu, A. (2001). New ex-

perimental apparatus for multimodal resonance imag-

ICPRAM 2020 - 9th International Conference on Pattern Recognition Applications and Methods

576

ing: initial epri and nmri experimental results. Physics

in Medicine & Biology, 46(4):1003.

Filippi, M., Preziosa, P., Banwell, B. L., Barkhof, F., Cic-

carelli, O., De Stefano, N., Geurts, J. J., Paul, F., Re-

ich, D. S., Toosy, A. T., et al. (2019). Assessment

of lesions on magnetic resonance imaging in multiple

sclerosis: practical guidelines. Brain, 142(7):1858–

1875.

Franchi, D., Gallo, P., Marsili, L., and Placidi, G. (2009). A

shape-based segmentation algorithm for x-ray digital

subtraction angiography images. Computer methods

and programs in biomedicine, 94(3):267–278.

Garc

ıa-Lorenzo, D., Francis, S., Narayanan, S., Arnold,

D. L., and Collins, D. L. (2013). Review of au-

tomatic segmentation methods of multiple sclerosis

white matter lesions on conventional magnetic reso-

nance imaging. Medical image analysis, 17(1):1–18.

Maurizi, A., Franchi, D., and Placidi, G. (2009). An op-

timized java based software package for biomedical

images and volumes processing. pages 219–222.

Placidi, G. (2012). MRI: essentials for innovative technolo-

gies. CRC Press.

Placidi, G., Cinque, L., Polsinelli, M., Splendiani, A., and

Tommasino, E. (2019). Automatic framework for

multiple sclerosis follow-up by magnetic resonance

imaging for reducing contrast agents. In International

Conference on Image Analysis and Processing, pages

367–378. Springer.

Placidi, G., Sabatini, M., Gallucci, M., and Sotgiu, A.

(2003). A calculation method for semi automatic fol-

low up of multiple sclerosis by magnetic resonance

eco planar perfusion imaging. Studies in Health Tech-

nology and Informatics, 95:164–169.

Roy, S., Butman, J. A., Reich, D. S., Calabresi, P. A., and

Pham, D. L. (2018). Multiple sclerosis lesion segmen-

tation from brain mri via fully convolutional neural

networks. arXiv preprint arXiv:1803.09172.

Steinman, L. (1996). Multiple sclerosis: a coordinated im-

munological attack against myelin in the central ner-

vous system. Cell, 85(3):299–302.

Thompson, A. J., Banwell, B. L., Barkhof, F., Carroll,

W. M., Coetzee, T., Comi, G., Correale, J., Fazekas,

F., Filippi, M., Freedman, M. S., et al. (2018). Di-

agnosis of multiple sclerosis: 2017 revisions of the

mcdonald criteria. The Lancet Neurology, 17(2):162–

173.

Vaidya, S., Chunduru, A., Muthuganapathy, R., and Krish-

namurthi, G. (2015). Longitudinal multiple sclero-

sis lesion segmentation using 3d convolutional neural

networks. Proceedings of the 2015 Longitudinal Mul-

tiple Sclerosis Lesion Segmentation Challenge, pages

1–2.

Valverde, S., Cabezas, M., Roura, E., Gonz

alez-Vill

S., Pareto, D., Vilanova, J. C., Rami

o-Torrent

a, L.,

Rovira,

A., Oliver, A., and Llad

o, X. (2017). Improv-

ing automated multiple sclerosis lesion segmentation

with a cascaded 3d convolutional neural network ap-

proach. NeuroImage, 155:159–168.

Yoo, Y., Brosch, T., Traboulsee, A., Li, D. K., and Tam, R.

(2014). Deep learning of image features from unla-

beled data for multiple sclerosis lesion segmentation.

In International Workshop on Machine Learning in

Medical Imaging, pages 117–124. Springer.

Guidelines for Effective Automatic Multiple Sclerosis Lesion Segmentation by Magnetic Resonance Imaging

577