Deep Learning for Computer-Aided Diagnosis (CAD) of Brain Diseases Four-Dimensional Magnetic Resonance Imaging (4D MRI Classification) from Glioma Tumor

Abstract

This work focuses on the utilization of Deep Learning and Convolutional Neural Networks (CNNs) for the accurate segmentation of a glioma tumor in brain magnetic resonance imaging (MRI) images. Specifically, we employed three distinct architectures using the labeled database BraTS2018 for training. To accomplish this, we utilized MATLAB 2019, which is compatible with Graphic Processing Unit (GPU) for enhanced performance. We have utilized various layers of a CNN to construct our architectures, including the convolutional layer, pooling layer, rectification layer, deconvolutional layer, and loss layer, to perform semantic segmentation. In addition, we compared the results obtained using the Dice coefficient. We show that the choice of epoch number has a great influence and a great importance for having the best results and giving better learning precision.

1. Introduction

The growing mass of data, often volumetric, attracts the intention of the practitioner, to motivate the design of new automatic methods for the analysis and interpretation of images. Manual segmentation is an essential operation for any treatment and diagnosis.

To do this, the doctor must know exactly the changes that have occurred in the images, but this is not always obvious because of several artifacts and other peculiarities related to the specificities of the object to be segmented (the anatomy of the brain) and the magnetic resonance imaging (MRI) acquisition process making segmentation a laborious and difficult process. Thus, it is necessary to have diagnostic tools.

These support systems allow radiologists to have accurate information about the characteristics of the regions of interest sought.

The goal of our work is to develop a tool for segmenting the sections of brain MRI images of glioma patients that makes up about a third of the most common primary brain tumors, based on deep learning, aimed at helping neuro-anatomists and neuroradiologists to identify the different forms of glioma lesions.

In this dissertation we develop three different convolutional neural network (CNN) models for glioma segmentation from four-dimensional magnetic resonance imaging (4D MRI) volumes, trying to increase the accuracy of the results by varying the parameters that make up the model.

2. Pathologies and brain tumors

A brain injury is a lesion that affects the brain. In general, it is a more or less extensive destruction of nerve tissue resulting in a deficit in perception, cognition, sensitivity, or motor skills depending on the role that the affected region played in the neurocognitive architecture.

After the age of 20, the human being loses thousands of neurons every day. This cellular degeneration is due to certain number of brain diseases such as brain tumors.

In our end-of-study project, we will focus on the detection of gliomas, which make up about a third of the most common primary brain tumors.

2.1 Glioma

Glioma is the most common form of a central nervous system (CNS) neoplasm that originates from glial cells in the brain [1]. These cells support the function of the other main type of brain cell: the neuron. Gliomas usually occur in the cerebral hemispheres of the brain, which are the largest and outermost parts of the brain that control many functions, including movement, speech, thinking, and emotions. However, they can also affect the brain stem, the lower part of the brain that controls functions such as breathing, blood pressure, heart rate, optic nerves, and the cerebellum, a part of the brain that deals with balance and other involuntary functions. These tumors can be benign or malignant. Generally, gliomas are classified into four grades. The fourth grade is the most aggressive type [2]: (1) Astrocytomas. (2). Ependymoma. (3). Glioblastoma. (4). Oligodendroglioma [3].

Headaches are the most common initial symptoms among glioma patients. These headaches are theoretically the result of tumor growth exerting a mass effect on the surrounding tissues. The mass effect, in turn, leads to pressure in the microvascularization and causes edema.

Depending on the tumor’s location in the brain, the mass effect can result in specific signs of a brain tumor. For instance, frontal lobe tumors may exhibit behavioral changes, while dominant temporal lobe tumors may cause receptive speech problems.

Other symptoms related to mass effects include nausea, vomiting, and vision changes. Seizures are the second most common presenting symptoms. The pathophysiology of seizures is attributed to tumor irritation of the cerebral cortex, leading to focal or generalized seizures. Additional symptoms with gliomas may involve tingling sensations, weakness, difficulty walking, and in rare cases, patients may present in a comatose state due to tumor hemorrhage, leading to acute hernia syndrome [1].

Like most primary brain tumors, the exact cause of gliomas is unknown. But certain factors can increase the risk of developing a brain tumor. These risk factors include: age; radiation exposure, and family history of glioma [3].

3. State of the art

Currently, several researchers are interested in the segmentation and classification of cerebral glioma by deep learning from MRI images. Among them we mention:

Mahmoud Khaled Abd-Ellah et al. [4], introduced a new network structure for the accurate detection and classification of gliomatous tumors, using two parallel deep convolutional neural networks (PDCNNs). The PDCNN structure comprises local, global, fusion, and output paths, which are employed to classify input images into normal or glioma/tumor images, and further distinguish between HGG (high-grade gliomas) and LGG (low-grade gliomas). The local and global paths are utilized to extract local and global characteristics, respectively, and each path consists of seven convolutional layers with ReLU activation and 7 Max Pooling. In the local path, the first convolutional layer utilizes a 5 × 5-pixel filter size to capture local characteristics, while the global path’s first convolutional layer uses a 12 × 12-pixel filter to extract global characteristics. After each convolutional layer in both paths, a Max Pooling is applied. Subsequently, the two paths are merged through a merging layer that establishes a cascading connection until the final output is reached. The merge path incorporates a normalization layer, followed by a rectified linear unit (ReLU) layer, and a fully connected layer connected to a dropout layer. The glioma classification process is facilitated on the output path using the Softmax function.

The performance of the proposed PDCNN was evaluated using the BraTS-2017 dataset, which comprises 600 normal two-dimensional (2D) images and 1200 abnormal 2D images. For the study, 1200 images were allocated to the training phase, 150 images to the validation phase, and 450 images to the test phase. The PDCNN structure achieved remarkable results in terms of accuracy, sensitivity, and specificity, with values of 97.44, 97.0, and 98.0%, respectively.”

Jakub Nalepa et al. [5], presented data augmentation techniques applied to MRI images of brain tumors in the BraTS-2018 database. These techniques improve the generalization capabilities of deep neural networks (DNNs) by increasing the size of training sets and can be perceived as implicit regularization. These augmentation algorithms can be divided into the following main categories: algorithms exploiting various transformations of the original data, including affine image transformations (rotation, zoom, recalibration, symmetry according to horizontal and vertical axes, or translations), elastic transformations, transformations at the pixel level (the modification of the brightness of the image, the application of gamma correction...), and various approaches to generating artificial data.

Thus, the results obtained showed that the increase of data was essential in the most efficient algorithms, and form much deeper and more complex neural networks, also to face the problem of limited ground truth data.

Kaldera et al. [6] developed a CNN-based process for localization and segmentation of glioma tumors from grayscale MRI (2D sections) images.

This dataset consists of 3064 T1-weighted images with enhanced contrast, collected from 233 patients with three types of brain tumors: meningioma (708 slices), glioma (1426 slices), and pituitary tumor (930 slices). In glioma segmentation process, they selected 123 axial MRIs of separate brain tumors for training and testing. This process is based on a faster region-based convolutional neural network (R-CNN) architecture that consists of two types of network: the region proposal network (RPN) and the region-based convolutional neural network (R-CNN). This technique is very popular in terms of transfer learning because it can train a classifier on a smaller dataset. The R-CNN acts as a classifier and its accuracy is based on the performance of the RPN algorithm. The latter is developed by adding additional convolutional layers that produce an objectivity score at various points in the image. It also produces the regional boundaries of these regions of interest. In addition, the analysis shows that the proposed technique gave average detection accuracy, sensitivity, dice coefficient, and confidence level of 99.81, 87.72, 91.14 and 93.6%, respectively.

For glioma segmentation, Vinay Rao et al. [7] applied deep neural networks (DNNs) to the entire MRI base of BRATS-2015 brain tumors to classify each pixel accordingly. For the glioma category, the dataset includes images from each modality (T1, T1c, T2, and Flair), where the model captures information related to the neighborhood of each pixel in all modalities and combines them to form a multimodal representation. In this work 32x32 patches in the XY, YZ, and XZ planes around each pixel for each modality are extracted. These patches are introduced to the DNN for each modality in order to learn good representations for each pixel.

Each DNN is formed separately to classify a pixel among non-tumor, necrotizing, edema, non-improving, and improving pixels. Each of the DNNs is formed as follows: two layers of convolution each are followed by a Max Pooling, then two fully connected layers followed by the RELU function and the Softmax function are used to produce the output. Afterward, they used concatenation of representations of all modalities as characteristics to form a random forest classifier to classify pixels. This method achieved an accuracy of 67% on a test dataset.

4. Theory of deep learning

Deep learning is a collection of machine learning techniques that has facilitated significant advancements in artificial intelligence (AI) in recent years. In machine learning (ML), a program analyzes a dataset to derive rules that can be used to make predictions about new data.

Deep learning is built upon artificial neural networks (ANNs), which are analogously composed of thousands of units (neurons), each performing small, simple operations. The outputs of one layer of neurons serve as inputs for the next layer, and this process continues through multiple layers. The progress in deep learning has been made possible, particularly due to the increase in computational power and the development of large databases (big data).

4.1 Convolutional neural networks

Convolutional neural networks (CNNs) are currently the most efficient models for the classification, localization and segmentation of images, especially in the medical field. The CNNs have two distinct parts. As input, an image is provided as a pixel matrix. It has two dimensions for a grayscale image. The color is represented by a third dimension of depth 3 to represent the fundamental colors [Red, Green, and Blue].

The first part of a CNN is the convolutional part itself. It works as an extractor of image characteristics. An image is passed through a succession of filters, or convolution cores, creating new images called convolution maps. Some intermediate filters reduce the resolution of the image by a local maximum operation. In the end, the convolution maps are flattened and concatenated into a characteristic vector, called CNN code.

This CNN code (Figure 1) at the output of the convolutional part is then connected as the input of a second part, consisting of fully connected layers (multilayer perceptron). The role of this part is to combine the characteristics of the CNN code to generate the output.

Figure 1.

Standard architecture of a CNN for image classification [8].

The output is the last layer with one neuron per category. The numerical values obtained are generally normalized between 0 and 1, of sum 1, to produce a probability distribution over the categories [8].

4.2 The most used deep learning segmentation models

4.2.1 Unet

U-Net is a model derived from the traditional convolutional neural network, specifically developed for biomedical image segmentation. It facilitates localization and segmentation by performing pixelwise classification, ensuring that the input and output share the same size. This renowned U-shaped model (Figure 2) is symmetrical and comprises two main parts:

Figure 2.

The architecture of the Unet model [9].

The left part, called the contraction path, consists of the general convolutional process, which involves multiple contraction blocks. Each block takes an input, applies two layers of 3x3 convolutions, followed by a RELU function, and Max Pooling (2x2). The number of filters or feature mappings doubles after each block, enabling the architecture to effectively learn complex structures [10].

The right part is the expansive path, consisting of transposed 2D convolutional layers. The number of expansion blocks matches the number of contraction blocks. Unet has the capability to perform precise image localization by predicting each pixel in the image individually [9].

4.2.2 SegNet

The SegNet model strikes a satisfactory balance between classification accuracy and computation time. Its symmetrical architecture allows for precise placement of abstract features in their correct spatial locations.

SegNet adopts an encoder-decoder architecture based on the convolutional layers of the VGG-16 model, which consists of 16 convolutional layers and is particularly attractive due to its uniform architecture [11]. The encoder is a series of convolutional layers followed by batch normalization (BN) and nonlinear activation functions (ReLU). Each block of two or three convolutions is subsequently followed by a subsampling layer (pooling) with a non-equal stride. The decoder is a mirror image of the encoder and possesses the same number of convolutions and blocks [12]. Figure 3 shows the architecture of the SegNet model:

Figure 3.

The segmentation architecture by SegNet [12].

5. Experimental results

Gliomas are brain tumors caused by the abnormal proliferation of glial cells in the central nervous system. These tumors make up about one-third of the most common primary brain tumors. Indeed, the interest in glioma detection has increased with the development of imaging and computing power. In this chapter, we focus on glioma segmentation from MRI images.

In this part, we will present the hardware resources, the software used as well as the database used. Then, we will present the different architectures proposed. Then we will make a comparative study between the different methods used to improve the performance of a model in terms of time or efficiency. We will conclude with a discussion of the results achieved.

5.1 The database

Several brain MRI databases have been published in the literature, we have chosen the BraTS database named Task01_BrainTumour that was carried out in 2018 [13], which contains three folders: imagesTr, imagesTs, and labelsTr. The BraTS dataset is a collection of MRI scans focused on brain tumors, particularly gliomas, which are the most prevalent primary brain malignancies. This dataset comprises 750 4-D volumes, with each volume representing a stack of three-dimensional (3D) images in nifti format (.nii). The size of each 4D volume is (240 x 240 x 155 x 4), where the first three dimensions correspond to the height, width, and depth of a 3D volumetric image. The fourth dimension corresponds to distinct scanning modalities, such as “FLAIR,” “T1w,” “t1gd,” and “T2w”. The dataset is divided into 484 training volumes with voxel labels and 266 unlabeled test volumes. Figure 4 shows a labeled 4D volume of the BraTS database.

Figure 4.

Examples of a labeled training volume.

5.2 Materials and methods

Several open source frameworks are available in the literature; we have chosen to work with the MATLAB R2019b language, which is easy to use.

Deep learning indeed requires substantial computational resources, and the availability of specialized resources like GPUs significantly impacts the user experience. Without these resources, the learning process can become excessively time-consuming, leading to longer training times and slower progress in learning from mistakes, which can be discouraging for users. Having access to powerful hardware accelerators like GPUs allows for faster training and more efficient learning, leading to a more positive and productive user experience. The experiments were all carried out on a machine that offers acceptable performance, the characteristics of which are as follows:

• Processor: AMD Ryzen threadripper 1950X 16-Core Processor 3.40 GHz.

• Installed memory (random access memory (RAM)): 16.0 GB.

• 1 TB hard drive.

• System type: Windows 10 operating system, 64-bit, x64 processor.

5.3 Methods developed

The segmentation of medical images plays a major role in image processing and diagnostic aid, especially in the case of brain diseases. Among these diseases, there are primary brain tumors, where gliomas are the most common and have a poor prognosis. So, in this work we present a technique of segmentation of gliomas from 4D MRI volumes using deep learning.

5.3.1 Preprocessing of data

The test volumes do not have labels, so we did not use the test data. Instead, we divided the 484 learning volumes into three independent sets that are used for learning, validation, and testing. Next, we preprocessed the training and validation data to train the CNN network more efficiently. The data preprocessing involves several essential operations:

• Cropping: The data are cropped to focus on the brain and tumor region, reducing the dataset size while preserving crucial information in each MRI volume and its corresponding labels.

• Normalization: Each modality of every volume is normalized independently by subtracting the mean and dividing by the standard deviation of the cropped brain region.

• Dataset split: The 484 learning volumes are divided into 400 for training, 29 for validation, and 55 for testing purposes.

• Random patch creation: Patches of size 132x132x132x4, containing both the image and corresponding pixel label data, are randomly generated. Each pair of volumes and labels contributes 16 randomly positioned patches to the training and validation sets.

• Validation data usage: The validation data are utilized to evaluate the network’s performance during the training process. It helps to determine whether the network is consistently learning, underlearning, or overlearning over time.

Next, we applied data augmentation operations to the training and validation data using:

• Random rotation of 90° and reflection of training data to make training more robust.

• Flipping horizontally and vertically.

• Cropping of response patches to network output size, 44 × 44 × 44 voxels.

5.3.2 Architecture of our network

In this part we proposed three different CNN architectures to perform binary semantic segmentation of brain tumors in MRI volumes, where each voxel is labeled as tumor or background.

5.3.2.1 First architecture (CNN 1)

As a first attempt, we started by forming a convolutional neural network of the most classic. This network consists of two convolutional layers each of which is followed by the ReLU activation function and a max pooling layer, then a deconvolutional layer is used to retrieve the size of the original input. Finally, the Softmax function is applied to produce the output. In this network, the voxel classification layer uses the Dice coefficient (allows the measurement of similarity in the labeled image and the resulting image) as the loss function to mitigate the problem of class imbalance in semantic segmentation problems.

The input volume, which has a size of 132x132x132x4, first moves to the first convolution layer. This layer is composed of 16 filters of size 3x3, followed by the RELU function that forces neurons to return positive values, and then we applied the Max Pooling to reduce the size of the image. This layer is composed of 16 filters of size 3x3, followed by the RELU function that forces neurons to return positive values, and then we applied the Max Pooling to reduce the size of the image. The second convolution layer consists of 32 filters of size 3x3, followed by a RELU and Max Pooling.

Then, we used a deconvolutional layer with seven size filters (2 × 2) in order to retrieve the original representation of our characteristics map and submit it to the last layer: Softmax.

The architecture of the first CNN formed is summarized in Figure 5. As learning options we have chosen: Optimiseur = Adam;MiniBatchSize = 1,Max Epochs = 10,InitialLearnRate = 5e-4, LearnRateSchedule = piecewise, LearnRateDropPeriod = 5,LearnRateDropFactor = 0.95,ValidationFrequency = 400,ExecutionEnvironment = GPU.

Figure 5.

Different architecture CNN formed. (a) The first CNN1; (b) the second CNN2; (c) the third CNN3 formed.

5.3.2.2 Second architecture (CNN2)

The architecture depicted in Figure 5 consists of three convolutional layers, each followed by the RELU function, a Max Pooling, and ultimately, a deconvolution layer. The Softmax function is then applied to generate the output. This network also uses the Dice coefficient as a loss function to classify voxels as tumors or backgrounds.

This time the input volume first passes through the first convolution layer. This layer is composed of 8 filters of size 3 × 3, followed by the RELU function and a Max Pooling of size (2 × 2). After applying the first convolutional layer with eight filters of size 3 × 3, the output of this layer is fed into the input of the second convolutional layer, which consists of 16 filters of size 3 × 3. The RELU activation function and a Max Pooling are then applied.

A third convolutional layer with 32 filters (3 × 3) was employed, followed by RELU activation and Max Pooling. Subsequently, a deconvolutional layer with 10 filters (2 × 2) was applied, and the architecture concluded with a Softmax layer.

In this case, we used the same CNN 1 learning options except the number of Epochs: Max Epochs = 50.

5.3.2.3 Third architecture (CNN3)

This third architecture, is composed of 4 convolutional layers with a different number of filters, each which is followed by a normalization layer (BN), the ReLU function and a Max pooling layer, then a deconvolution layer is used. Finally, the Softmax function is applied to produce the output. This network also uses the Dice coefficient as a loss function.

This time the input volume passes through four convolution blocks consisting of the following:

• The first block contains a convolution layer with eight filters of size 3x3, followed by BN, ReLU, and a Max pooling of size (2 × 2). The result of this block is introduced at the entrance of the second which contains a convolution layer with 16 filters of size 3x3, followed by BN, ReLU, and a Max pooling of size (2 × 2). Then, we used a third block with a convolutional layer consisting of 32 filters of size 3 × 3, followed by BN, ReLU and a Max pooling of size (2 × 2).

• The last block contains a convolution layer with 64 filters of size 3x3, followed by BN, ReLU, and a Max pooling of size (2 × 2).

Then, we applied a deconvolutional layer with 13 size filters (2 × 2). And the last layer was a Softmax. In this architecture, we used the same learning options of CNN 1.

5.4 Outcomes and discussion

During our experiments, we created three architectures where we applied the learning options by training the network using Adam’s optimization solver and specifying the parameters using the “Training Options” function. These options are used to monitor the progress of the network’s training. The initial learning rate is set at 5e-4, and then, it is adjusted regularly according to the MiniBatchSize.

In order to show the results obtained for these architectures, we illustrate in the following the results in terms of accuracy and error for each of the architectures.

5.4.1 First CNN

During training, the processor took 140.8 min to do 8000 iterations (finish training) with an error rate of 0.000475 and an accuracy of 99.08%.

Figure 6 shows the training progress of the CNN1 where the blue plot represents the training accuracy and the red plot represents the training loss.

Figure 6.

Final progression of learning CNN1.

Figure 7 shows the result of the semantic segmentation of gliomas by CNN1, where the tumors appear in red, the brain in gray, and the background in blue.

Figure 7.

Segmentation result using CNN1; (a) field truth image; (b) predidated image.

5.4.2 Second CNN

We launched a learning of the second CNN. The processor took 884.11 min for 40,000 iterations with an error rate of 0.00031512 and an accuracy of 98.08%.

Figure 8 shows the training progress of the CN2.

Figure 8.

Final progression of CNN2 learning.

Figure 9 below shows the result of the semantic segmentation of gliomas using CNN2.

Figure 9.

Segmentation result using CNN2; (a) field truth image; (b) predidated image.

5.4.3 Third CNN

Learning the third CNN took 314.03 minutes to complete 8000 iterations with an error rate of 0.000475 and an accuracy of 99.03%. Figure 10 shows the training progress of the CNN3.

Figure 10.

Final progression of CNN3 learning.

Figure 11 below shows the result of semantic segmentation of gliomas using CNN3 from 4D MRI volumes.

Figure 11.

Segmentation result using CNN3; (a) field truth image; (b) predidated image.

5.4.4 Dice coefficient

In this work, we tested the performance of the proposed CNN models using 55 image 4D MRI volumes from our database.

In order to calculate the accuracy of the segmentation between the field truth images and the images predicted by the different architectures, we calculated the Dice coefficient.

The Dice coefficient or the coefficient of similarity indicates the positive correlation between two images, it varies between 0 and 1, where 1 means the greatest similarity between prediction and truth.

Table 1 below shows the comparison between the different architectures used for glioma segmentation in terms of the Dice coefficient.

Architecture Coefficient de Dice	CNN1	CNN2	CNN3
For the substance	0.99834	0.99821	0.99797
For the tumor	0.83898	0.84799	0.79892

Table 1.

Coefficient de Dice obtenu par les différentes architectures.

The results obtained by the three architectures are respectively, 0.99834, 0.99821, and 0.99797 for the background, and 0.83898, 0.84799, and 0.79892 for tumors.

Thus, the results obtained are satisfactory and encouraging. The performance of the model can be enhanced by employing deeper and more complex architectures. Additionally, increasing the number of epochs and adjusting the MiniBatchSize, along with modifying the learning options, can also lead to improved results. These adjustments allow the model to learn more effectively and extract more intricate features from the data, ultimately enhancing its overall performance.

5.4.5 Comparison

After analyzing the results obtained, we note the following remarks:

Based on Figures 7,9 and 11 information was previously featured in the CNN1, CNN2, and CNN3 architectures. The accuracy of learning generally improves with an increase in the number of epochs. This indicates that with each epoch, the model gains more information and refines its understanding of the data. If the accuracy decreases, it implies that the model requires more information to learn effectively. In such cases, increasing the number of epochs would be beneficial as it allows the model to accumulate more knowledge and enhance its performance.

Furthermore, as the number of epochs increases, the learning error tends to decrease. This reduction in error signifies that the model is converging toward a better solution and is becoming more proficient at the given task with each epoch. Therefore, training the model for more epochs can lead to better generalization and improved performance. So we find that in our work, architecture 2 is better than architecture 1 and architecture 3.

Also, the deeper the architecture, the better and more accurate the test result. Absolutely, the number of epochs and MiniBatchSize are crucial parameters in the learning process, and finding the right balance is essential for model performance.

As you rightly mentioned, the number of epochs allows the model to learn from the data more effectively, avoiding issues of overlearning or underlearning. Adequate epochs enable the model to converge to an optimal solution while preventing it from memorizing the training data or failing to capture its underlying patterns.

Regarding MiniBatchSize, while larger values can offer benefits in terms of faster training and improved generalization, working with a MiniBatchSize of 1 (online learning) is a valid approach, especially when hardware limitations prevent increasing the size. Although this might slow down the training process, stochastic gradient descent (MiniBatchSize of 1) can still optimize the model’s parameters and yield reasonable results.

In situations where you are constrained by hardware limitations, it is essential to explore other techniques for model optimization. You can consider learning rate schedules, early stopping, and regularization to improve convergence and prevent overfitting. Additionally, exploring more computationally efficient model architectures can help work within the hardware capacity while achieving meaningful outcomes.

Remember, in practical applications, the balance between available resources and desired results often plays a crucial role, and many successful models can be built and trained with limited hardware settings. Adaptability and innovation in model design and optimization are key to obtaining satisfactory performance under such constraints.

6. Conclusion

The segmentation of medical images remains a very broad field of research. In the context of brain imaging, the goals of segmentation of brain MRI images are indeed multiple: help with diagnosis, monitoring of the evolution of the patient’s condition, clinical test… etc.

The focus of our work is on the semantic segmentation of gliomas which are among the most common primary brain tumors, from magnetic resonance imaging (MRI) using convolutional neural networks (CNNs) that have shown its performance in recent years.

We have developed in this end-of-study project three different CNN architectures for the segmentation of gliomas from 4D MRI volumes inspired by an example of segmentation provided by MATLAB 2019. In the training phase, the use of a CPU makes the execution time too expensive. In order to solve this problem, it is necessary to use deep convolutional neural networks deployed on a GPU instead of a CPU.

To carry out our segmentation work, we used the BraTS database carried out in 2018, to learn the models and carried out the validations and testing part.

Network parameters are difficult to define a priori. That’s why we have defined different models with different architectures to achieve better results in terms of accuracy and error.

The results found are satisfactory (a Dice coefficient for tumors of 83.8, 84.7 and 79.89% for the three architectures CNN1, CNN2 and CNN3, respectively), which allowed us to say that the use of semantic segmentation methods (deep learning) allows us to give better segmentation results.

Several perspectives can be envisaged in the extension of this dissertation, we can quote:

• Test other larger databases, and use deeper architectures like Unet and SegNet.

• Develop a tumor segmentation architecture based on 2D images.