Automated Brain Lobe Segmentation and Feature Extraction from

Multiple Sclerosis Lesions Using Deep Learning

Nada Haj Messaoud

1,2 a

, Rim Ayari

Asma Ben Abdallah

and Mohamed Hedi Bedoui

Faculty of Sciences of Monastir (FSM), University of Monastir, Monastir, Tunisia

Medical Technology and Image Processing Laboratory, Faculty of medicine, University of Monastir, Monastir, Tunisia

Keywords: Brain Lobes Segmentation, Deep Learning, Multiple Sclerosis Lesion, U-Net, Features Extraction.

Abstract: This study focuses on automating the segmentation of brain lobes in MRI images of Multiple Sclerosis (MS)

lesions to extract crucial features for predicting disability levels. Extracting significant features from MRI

images of MS lesions is indeed a complex task due to the variability in lesion characteristics and the detailed

nature of MRI images. Furthermore, all these studies required continuous patient monitoring. Therefore, our

contribution lies in proposing an approach for the automatic segmentation of brain lobes and the extraction of

lesion features (number, size, location, etc.) to predict disability levels in MS patients. To achieve this, we

introduced a model inspired by U-Net to perform the segmentation of different brain lobes, aiming to

accurately locate the MS lesions. We utilized two private and public databases and achieved an average mean

IoU score of 0.70, which can be considered encouraging. Following the segmentation phase, approximately

7200 features were extracted from the MRI scans of MS patients.

1 INTRODUCTION

Multiple sclerosis (MS) is a demyelinating disease of

the central nervous system (CNS) characterized by

damage to the protective myelin surrounding the nerve

fibers within the brain and spinal cord. It primarily

affects young adults and leads to increasing disability

(Thompson, et al., 2018). Diagnosis is confirmed

through magnetic resonance imaging (MRI), with

varying contrast in cerebral MRI. MS lesions are

surrounded by edema, which appears as a

hyperintense signal on the T2 FLAIR image. These

lesions can appear in different areas of the brain. They

are characterized by their variability in terms of

volume, location, shape, subjects, and texture, leading

to symptoms that vary depending on where these

lesions are located. Consequently, the cerebral lobes

are also vulnerable to the impact of MS, as they

contain numerous nerve fibers and play a crucial role

in various brain functions. So, MS Lesion appears in:

• The temporal lobe can affect vision, touch,

memory, hearing, and language comprehension.

http://orcid.org/0000-0001-6243-1373

http://orcid.org/0000-0002-8292-7656

http://orcid.org/0000-0001-7821-7734

http://orcid.org/0000-0003-4846-1722

• The frontal lobe can lead to issues with

emotional control, cognitive functions, planning,

decision-making, as well as the supervision of

voluntary movements and activities.

• The parietal lobe can disrupt the processing of

information related to temperature, taste, touch, and

movement.

• The occipital lobe can lead to vision problems,

such as visual perception alterations, visual

disturbances, and even partial or total vision loss.

Thus, extracting meaningful features from brain

lesions to classify these anomalies based on cerebral

lobes can provide valuable insights into predicting

which human activities or tasks may be affected by

these abnormalities. Therefore, to extract these

features, a step of segmenting the different cerebral

lobes is required to facilitate the localization of brain

lesions. However, automatic brain region

segmentation is challenging due to variations of brain

size and shape from one individual to another, as well

as variations in the quality, size, and number of MRI

slices. Furthermore, cerebral lobe segmentation is

532

Haj Messaoud, N., Ayari, R., Ben Abdallah, A. and Bedoui, M.

Automated Brain Lobe Segmentation and Feature Extraction from Multiple Sclerosis Lesions Using Deep Learning.

DOI: 10.5220/0012390700003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 3: VISAPP, pages

532-540

ISBN: 978-989-758-679-8; ISSN: 2184-4321

typically performed using 3D data, making the use of

2D slices with a reduced number of MRI slices, not

an easy task. Indeed, collecting data on multiple

sclerosis (MS) can be a challenging task, especially

when it involves both imaging and clinical data. To

meet our research needs and consider our medical

context, we utilized two separate databases: our own

database and the public database proposed by

(Almutairi, 2022). The characteristics of the two

databases are similar, that’s why we chose to combine

them to perform the segmentation task.

For the segmentation of brain regions, several

tools and methods based on machine learning (ML)

and Deep Learning (DL) have been proposed, each

having its own advantages and disadvantages. In this

research, our approach involves utilizing the U-Net

architecture to segment the distinct brain lobes. This

segmentation aims to facilitate the extraction of

features from MS lesions, pinpoint their locations,

and predict the disease's progression. The objective of

the study is the segmentation of these regions using

only 2D data with a limited number of slices.

The structure of this paper is as follows. Related

works are detailed in Section 2. Datasets are

explained in Section 3. The proposed workflow is

detailed in Section 4. Section 5 provides the results of

the method we proposed. Finally, Sections 6 closes

with a discussion, conclusion and main limitations of

this study.

2 RELATED WORKS

Accurate segmentation of MRI images becomes a

crucial task for Alzheimer's disease, dementia, partial

epilepsy, multiple sclerosis, etc. It has become an

important task for many evaluations in neurological

research, including the diagnosis, progression and

treatment of various neurological diseases such as

neurological diseases. Manual segmentation is

considered the gold standard in the field of

anatomical segmentation. Determining brain

structure is time-consuming and detailed because an

MRI can consist of hundreds of segments, depending

on the resolution of the MRI. Therefore, this labor-

intensive method is not suitable for large-scale

neuroimaging studies. Automatic segmentation

techniques attempt to resolve the limitations

associated with manual segmentation.

Following the advancement of algorithms and

computing resources several segmentation techniques

have been developed. These techniques include (a)

FreeSurfer, (b) FMRIB Software Library (FSL), and

2021)… Moreover, approaches that utilize

convolutional neural networks (CNN) are widely

employed for automatic segmentation tasks.

Table 1: State of the art of published works on the

segmentation of brain structures.

Authors Model Application Description

(Sken,

2016)

FCN

with a

late

fusion

method

Tissue

segmentation

FCNs for the

segmentation

of isointense

phase brain

MR images.

(Klein,

2018)

DeepNA

Anatomical

segmentation

3D Deep

convolutional

neural

network for

the automatic

segmentation

NeuroAnaTo

(Ayed,

2018)

3D CNN

Anatomical

segmentation

3D and fully

convolutional

neural

network

(CNN) for

subcortical

brain structure

segmentation

in MRI.

(Llado,

2019)

FCNN

Tissue

segmentation

Eight FCNN

architectures

inspired by

robust stateof-

the-art

methods on

brain

segmentation

related tasks.

(Tang,

2019)

U-net

Tissue

segmentation

Skip-

connection U-

net for WM

hyper

intensities

mentation.

(Gao,

2020)

Fuzzy C-

mean

Tissue

segmentation

Fuzzy C-

means

framework to

brain tissue

segmentation

The Table 1 summarizes the different automated

segmentation techniques mentioned in the literature.

FreeSurfer is an open-source software. Its focus is

on processing 3D images, including full volumetric

MR images. It is optimized for segmentation, 3D

brain reconstruction, and volumetric analysis from

Automated Brain Lobe Segmentation and Feature Extraction from Multiple Sclerosis Lesions Using Deep Learning

533

these types of images. FSL is a comprehensive library

of neuroimaging tools for structural, functional, and

diffusion tensor imaging (DTI) studies. It is capable

of handling a wide range of data and is known for its

robustness. SPM is a package developed for the

analysis of neuroimaging data coming from several

imaging modalities. It is capable of performing

complex and detailed statistical analyses, but requires

MATLAB platform, volBrain (Online Web Platform)

is a web-based pipeline for MRI brain volumetry. Its

system is primarily based on a multi-atlas, patch-

based segmentation method. Registration is necessary

to use this platform, but there is a limit on the number

of concurrent jobs that can be submitted. Although it

is user-friendly, it does not offer as much as more

advanced software, which results in limitations of

specific tasks. Although there are a variety of tools to

analyze functional and structural imaging, they are

not specifically designed for processing 2D data with

fewer slices. In addition to the segmentation tools

previously presented, there are several works

proposed for the segmentation of brain regions based

on deep learning. The Table 1 provides a summary of

some model. The majority of previous studies have

typically aimed to segment various brain tissues,

including white matter, gray matter, and

cerebrospinal fluid. However, our approach is

distinct. Our research focuses on the segmentation of

brain lobes, which allows for the precise localization

of MS lesions. By concentrating on this specific

region of the brain, we can thoroughly evaluate how

these lesions impact human cognitive and sensory

functions, offering a unique perspective on the

consequences of MS.

3 CONSIDERED DATASETS

Two datasets were used: our private MS database and

the public database proposed by (Almutairi,

2022),(https://data.mendeley.com/datasets/8bctsm8

jz7/1).

3.1 MS Private Dataset

We recruited 22 patients diagnosed with relapsing-

remitting multiple sclerosis (RR-MS) and obtained

informed consent from the Fattouma Bourguiba

University Hospital Ethics Committee. MRI T2-

FLAIR image sequences were acquired at the

Neurology Department of Fattouma Bourguiba

University Hospital in Monastir. Each patient had

multiple time-points, ranging from 2 to 4. The dataset

was generated using a Philips 1.5T machine (Ingenia,

Philips, medical systems, Best, the Netherlands)

equipped with a 20-channel phased-array coil for the

head, neck, and spine, located at the Fattouma

Bourguiba Hospital Medical Imaging Department.

The original images of MS patients varied in

dimensions from (352 × 352) to (512 × 512), with a

spatial resolution of (0.46 × 0.46) and a slice

thickness of 7 mm. Ground truths were meticulously

prepared and validated by our highly experienced

expert with 16 years of expertise (Figure 2). Among

these 22 patients, clinical data is accessible for 19 of

them. Demographic and lesion characteristics are

detailed in Table 2 and Table 3.

Table 2: Baseline characteristics of MRI Image (Private

dataset).

Baseline characteristics (MRI Ima

Sex ratio

(

male : female

)

0,46 [7:15]

Modalit

T2- FLAIR

Image Size 256×256

Number of original images 370

Number of groundtruth 370

Total number of lesions

(SD) [min : max]

1716 ± 3.98 [1:26]

Surface area of lesion (mm

)

[min : max]

(

)

[4.0, 859.0] ± 93.63

Table 3: Baseline characteristics of Clinical Data (Private

dataset).

Baseline characteristics (Clinical Data)

Sex ratio

(

male : female

)

0,35 [5:14]

Current age (years) [min :

max]

35 ± 9.2 [20 : 50]

e of onset 8,05 ± 3,48 [3 : 16]

EDSS [min : max] 3,10 ± 2,35 [1 : 8]

Types of Medicines [1,2,3,4]

Co-moroidity (No/Yes) (16/3)

3.2 MS Public Dataset

Offered by (Almutairi, 2022) in 2022, this dataset

constitutes a valuable resource in the field of multiple

sclerosis (MS) (see Table 4 and Table 5). It comprises

multi-sequence MRI (1.5 Tesla) data from 60 patients

diagnosed with MS, accompanied by a consensus-

based manual segmentation of lesions, assessments of

disability levels using the Expanded Disability Status

Scale (EDSS), general patient information, and

relevant clinical data. One of the dataset's standout

features is the quality of its manual lesion

segmentation. Two expert radiologists and a

neurologist, ensuring a high level of accuracy and

reliability, performed this segmentation. It covers

three crucial MRI sequences: T1-weighted, T2-

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

534

weighted, and FLAIR (fluid-attenuated inversion

recovery), enabling in-depth analysis of lesion

characteristics in various contexts.

Table 4: Baseline characteristics of MRI Image (Prublic

dataset).

Baseline characteristics (MRI Images)

Modalities T1/T2/T2-

FLAIR

Number of OI

FLAIR 1446

Number of GT

(FLAIR) with at

least one lesion

794

Number of OI

(

)

1385

Number of GT

(T2) with at least

one lesion

644

Number of OI (T1) 1358

Table 5: Baseline characteristics of Clinical Data (Public

dataset).

Baseline characteristics (Clinical Data)

Sex ratio (male : female) 0,30 [14:46]

Current age (years) [min :

max]

34 ± 12,1 [15 : 56]

e of onset 29,7 [8 : 52]

EDSS [min : max] 2,26 ± 1,5 [0 : 6]

Types of Medicines [1,2,3,4,5]

Co-moroidity (No/Yes) (47/13)

4 PROPOSED WORKFLOW

Our proposed pipline is composed of six steps (Figure

1): (i) Pre-processing to refine images and facilitate

brain lobes segmentation, (ii) Clinical data

preprocessing to ensure their compatibility with AI

algorithms, achieving greater consistency. (iii) Brain

lobes Segmentation based on DL architecture, (iv)

Automated feature extraction for capturing

significant characteristics including lesion Number,

lesion size, localization, and lobe area... (v) Creating

a file that combines both MRI features and clinical

data to identify the most correlated features to the MS

patients' disability levels and (vi) EDSS Prediction.

Each of these steps will be presented in the following

section.

4.1 Pre-Processing

In this phase, we applied preprocessing operations to

both the image data and clinical data as follows:

Original Image (OI)

Groundtruth (GT)

Preprocessing of Image Data

• Skull Stripping: We isolated the region of

interest by extracting it from extracranial and

non-cerebral tissues.

• Background Reduction: The black background

was minimized through cropping operations

• Image Resizing: All MRI images were resized

to (256× 256) in order to standardize the

database.

Preprocessing of Clinical Data

• Z-Score Normalization: We used the Z-Score

method to normalize variables such as age and age

of onset. The new value is calculated using the

formula (x - μ) / σ, where x represents the original

value, μ is the mean of the data, and σ is the

standard deviation of the data.

• Encoding: We applied encoding to variables such

as gender, comorbidities, presenting symptoms,

and type of medicines.

4.2 Brain Lobes Segmentation

Automatically segmenting brain lobes presents a

complex set of challenges due to the inherent

variability in brain shapes, sizes, and abnormalities,

as well as the diverse qualities and sizes of brain MRI

scans. Traditional methods rely on 3D MRI data,

making segmentation with 2D scans particularly

complex (Singh, 2021).

Multiclass segmentation is an advanced computer

vision task that extends beyond binary segmentation.

It categorizes every pixel in an image into distinct

classes, allowing for the differentiation of various

anatomical structures in medical imaging within an

MRI scan, such as the brain, heart, and lungs. For our

initial work, we chose to utilize the U-Net

(Ronneberger, Fischer, & Brox, 2015) architecture, as

it continues to be a reference in the field of medical

image segmentation (Liang, 2018) (Wen, 2019). This

choice was based on its proven effectiveness and

reliability in accurately segmenting medical images.

It features an encoding and decoding path, which

progressively reduces and increases spatial

resolution. U-Net can simultaneously segment

multiple classes in a single pass, with each class

corresponding to a specific object or region. This

capability is crucial for tasks like semantic

segmentation, where precise classification of

different categories within an image is essential.

Automated Brain Lobe Segmentation and Feature Extraction from Multiple Sclerosis Lesions Using Deep Learning

535

Figure 1: Workflow proposed for Brain Lobes Segmentation and MS lesions features extraction.

U-Net's skip connections maintain spatial

information, ensuring accurate segmentation.

Consequently, it is a fundamental tool in medical

image analysis. To achieve this, we began by

preparing the ground truth data for each cerebral lobe

in collaboration with our expert. We then proceeded

to implement the U-Net model. In total, we had seven

classes to segment, which included the Frontal Lobe

(FL), Occipital Lobe (OL), Parietal Lobe (PL),

Temporal Lobe (TL), Brain Stem (BS), Cerebellum

(C), and the background (B). Figure 2 provides an

example of the prepared ground truth data. Figure 3

depicts the functioning of each cerebral lobe,

highlighting the specific regions and functions of

each lobe. This illustration aids in gaining a better

understanding of how different parts of the brain

interact to control various aspects of cognition,

perception, and movement. It serves as a valuable

resource for studying MS as it helps us comprehend

the potential impacts of lesions on different brain

functions.

Figure 2: Example of Brain Lobe groundtruth.

U-Net wass proposed by (Ronneberger et al. 2015) in

2015 for biomedical image segmentation. It consisted

of a contraction path (downsampling) associated with

an expansion path (upsampling). It was proposed to

overcome the major limitation of the traditional CNN,

which is a compromise between location accuracy,

represented by low-level features, and contextual

information, provided by higher-level features.

During the contraction path, spatial information is

reduced while feature information is increased.

Figure 3: Illustration of the functioning of each cerebral

lobe.

However, during expansion, upsampling is

performed through the transposed convolutions to

build the segmented image. It is characterised by skip

connections (concatenation) between these two paths

for more accurate retrieval of spatial information. We

operated two modifications : (i) Batch Normalisation

to speed up learning and produce accurate models. (ii)

Dropout between the two consecutive convolutional

layers to avoid overfitting (Ronneberger, Fischer, &

Brox, 2015). Table 6 sums up the hyper-parameters

of convolution and deconvolution layers used in this

model. The following notations are used: BN stands

for Batch Normalisation and ReLu stands for

Rectified Linear Unit.

For the implementation of U-

Net model, Table 5 sums up the hyper-parameters of

convolution and deconvolution layers used in this

model.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

536

Table 6: U-Net structure details for the two paths.

Type Taille/Nombre of filters Padding Stride kernel_initializer Parameters

Convolution

((3 × 3 × 16) + BN +ReLu) ×2 + Dropout (0.1)

he_normal

1,946,807

((3 × 3 × 32) + BN +ReLu) ×2 + Dropout (0.1)

((3 × 3 × 64) + BN +ReLu) ×2 + Dropout (0.1)

((3 × 3 × 128) + BN +ReLu) ×2 + Dropout (0.1)

((3 × 3 × 256) + BN +ReLu) ×2 + Dropout (0.1)

MaxPooling (2x2)

Conv2DTranspose (2x2) 2

4.3 Implementation Details

The training phase requires establishing a set of

parameters such as the optimiser, learning rate,

number of epochs, and batch size . . . These are

usually experimentally selected or based on recent

studies with the aim of producing precise

segmentation performance. The implementation is

conducted onIntel Core i9-11900F @ 2.50 GHz,

32Go RAM and a Nvidia GeForce RTX 3090. The

suggested models were implemented in Python

language using Keras with Tensorflow backend.

Table 7 summarizes the hyperparameters used.

Table 7: hyper parameter used in the training step.

Optimizer Learning rate Batch_Size Epochs

ADAM 0.001 16 200

As metrics, we used accuracy, recall, and dice to

evaluate the model during training.

𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦  𝑇𝑁  𝑇𝑃/𝑇𝑃  𝐹𝑃  𝑇𝑁  𝐹𝑁 (1)

𝐷𝑖𝑐𝑒  2 𝑇𝑃/2𝑇𝑃  𝐹𝑃  𝐹𝑁

(2)

𝑅𝑒𝑐𝑎𝑙𝑙  𝑇𝑃/𝑇𝑃  𝐹𝑁

(3)

4.4 Feature Extraction Process

McDonalds is considered standard diagnostic criteria

for multiple sclerosis (MS) based on clinical,

radiological and other medical data. These criteria

allow doctors to confirm the diagnosis based on

several pieces of evidence such as the patient's

symptoms, the results of magnetic resonance imaging

(MRI) examinations and other tests. They are used to

standardize the diagnostic process and to guarantee

consistency. Consistent with these criteria

(Skripuletz, 2019), we extracted FLAIR MRI features

and specific disease characteristics based on feature

types, including lesion location, shape, size, number,

and density (Figure 4). To do this, a graphical

interface has been developed.

Figure 4: Example of Features extracted from MRI Images.

The main goal of this interface is to make the feature

extraction process more accessible to non-technical

users or experts in a particular field. This helps speed

up the data analysis process, especially when dealing

with large amounts of information or complex data.

The proposed extraction steps are as follows:

• Measure the centroid of the ventricle to identify

lesions located near the cerebral ventricular

system.

• Load the brain lobes and calculate the area of each

lobe.

• Segment MS lesions and calculate the area,

location and number of each lesion.

• Create a file containing all the features extracted

from FLAIR MRIs.

• Integrate clinical data with imaging data to

facilitate predictive modeling, ultimately

identifying features with the strongest correlations

with disability progression.

Figure 5 presents a visual representation of the

process we proposed. It allows you to better

understand how our methodology works.

For the MS lesion segmentation phase we used

our own "Concat-U-Net" method published in

(Messaoud, 2022) which makes it possible to segment

objects of variable size, location and number such is

the case of lesions.

Automated Brain Lobe Segmentation and Feature Extraction from Multiple Sclerosis Lesions Using Deep Learning

537

Figure 5: Workflow proposed for the MS lesion features

extraction.

5 RESULTS OF BRAIN LOBE

SEGMENTATION

In total, we prepared the ground truths of 32 patients.

We implemented a nested 5-fold cross-validation

over the whole datasets. The training curves of the U-

Net model are presented in Figure 6 and Figure 7. The

letters S stand for Subject, and the numbers represent

the subject identifier. (– denotes the subjects from 1

to N). Using the previously presented 5-Fold cross-

validation scheme, we have successfully applied

Deep learning to segment the cerebral lobes, even

with a limited number of slices for each patient. This

approach has significantly improved our ability to

accurately locate multiple sclerosis (MS) lesions and

estimate their size.

In the works of (Almutairi, 2022), the approach

involved segmenting the various brain lobes through

a series of steps. These steps included dividing the

brain into four subregions, measuring the center,

width, and height of each region, and subsequently

segmenting each subregion into four additional

sections, each of which was labeled accordingly.

However, in our case, we utilized a U-Net-inspired

architecture to calculate the surface of each cerebral

lobe, enabling us to extract the percentage of lesion

involvement in the brain lobes. This information

could serve as a significant biomarker for multiple

sclerosis diagnostic analysis.

Figure 6: Accuracy, Recall, Dice curves.

Figure 7: Training loss and Validation Loss curves.

To evaluate the obtained result, we used the Mean

Intersection over Union (MeanIoU) is a metric used

to assess the accuracy of image segmentation models.

It calculates the intersection over the union for each

class and then computes the average across all classes.

The formula for calculating MeanIoU is as follows:

MeanIoU 







∑







∩







∪









(4)

The boxplots presented in Figure 8 depict the

results for each testing level. As an average result, we

achieved a mean IoU (Intersection over Union) score

of 0.70, which can be deemed highly encouraging.

Figure 8: Boxplots showing the performance of tested

model with all results obtained.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

538

Figure 9: Example of Output segmentation results.

This high score indicates a strong match between the

model's predictions and the ground truths,

underscoring the effectiveness of our cerebral lobe

segmentation approach for assessing multiple

sclerosis lesions. This can be readily appreciated in

Figure 9, which is also reflected in the measured

MeanIoU value of the implemented U-Net. After

performing the segmentation of brain regions, we

followed the workflow presented in Figure 5 to

extract features from MS lesions. These features were

then integrated with clinical and demographic data to

study the correlation between the features and their

impact on the progression of this pathology. In total,

we extracted +7200 features. The calculation of these

features was performed using the following formula:

Total number of features = (Number of features per

slice) x (Number of slices per patient) x (Number of

patients). The proposed works is published in

https://github.com/nadandan/MRI-Brain-Region-Seg

mentation

6 DISCUSSION AND

CONCLUSION

The key contribution of this study is the segmentation

of brain lobe regions using only 2D data with a

limited number of slices and proposing an automated

approach to extract features from MS lesions and

combine them with patients' clinical and

demographic data. As our initial step, we selected the

U-Net architecture, which has demonstrated superior

performance in medical image segmentation. We

developed a U-Net-inspired model for the

segmentation of seven classes: the Frontal Lobe (FL),

Occipital Lobe (OL), Parietal Lobe (PL), Temporal

Lobe (TL), Brain Stem (BS), Cerebellum (C), and the

background (B). This was accomplished using 2D

data with a reduced number of slices for each patient.

On average, we achieved a highly encouraging mean

Intersection over Union (IoU) score of 0.70. Our

focus lies in examining the surface, localization, and

the number of lesions. Consequently, we successfully

extracted approximately 7200 features. In our future

work, we intend to utilize the generated feature file

(both clinical and image data) from our interface to

predict the disability level of MS patients. This

study's limitation lies in its exclusive use of the U-Net

architecture. It is essential for us to assess other

architectures for performance comparison and

possibly develop our unique model. Furthermore,

expanding the dataset with more images is crucial for

improving overall performance.

REFERENCES

Almutairi, A. M. (2022). Brain MRI dataset of multiple

sclerosis with consensus manual lesion segmentation

and patient meta information. Data in Brief.

Ayed, J. D. (2018). 3D fully convolutional networks for

subcortical segmentation in MRI: A large-scale study.

NeuroImage.

Gao, P. L. (2020). Temporally Consistent Segmentation of

Brain Tissue From Longitudinal {MR} Data. {IEEE}

Access.

Klein, C. W. (2018). DeepNAT: Deep convolutional neural

network for segmenting neuroanatomy. NeuroImage.

Automated Brain Lobe Segmentation and Feature Extraction from Multiple Sclerosis Lesions Using Deep Learning

539

Liang, Z. Z. (2018). Unet++: A nested u-net architecture

for medical image segmentation. Lecture Notes in

Computer Science (including subseries Lecture Notes

in Artificial Intelligence and Lecture Notes in

Bioinformatics).

Llado, J. B. (2019). Quantitative Analysis of Patch-Based

Fully Convolutional Neural Networks for Tissue

Segmentation on Brain Magnetic Resonance Imaging.

IEEE Access.

Messaoud, N. H. (2022). Automated segmentation of

multiple sclerosis lesions based on convolutional

neural networks. Computer Methods in Biomechanics

and Biomedical Engineering: Imaging and

Visualization.

Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net:

Convolutional networks for biomedical image

segmentation. International Conference on Medical

image computing and computer-assisted intervention.

Singh, M. K. (2021). A Review of Publicly Available

Automatic Brain Segmentation Methodologies,

Machine Learning Models, Recent Advancements, and

Their Comparison. Annals of Neurosciences.

Sken, D. N. (2016). Fully convolutional networks for multi-

modality isointense infant brain image segmentation.

International Symposium on Biomedical Imaging.

Skripuletz, P. S.-W. (2019). Impact of the {McDonald}

Criteria 2017 on Early Diagnosis of Relapsing-

Remitting Multiple Sclerosis. Frontiers in Neurology.

Tang, J. W. (2019). Skip connection U-Net for white matter

hyperintensities segmentation from MRI. IEEE Access.

Thompson, A. (2018). Diagnosis of multiple sclerosis: 2017

revisions of the McDonald criteria. The Lancet

Neurology

Wen, F. C. (2019). An Improved Framework Called Du++

Applied to Brain Tumor Segmentation. 15th

International Computer Conference on Wavelet Active

Media Technology and Information Processing,

ICCWAMTIP 2018.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

540