Intelligence-Based Medicine 11 (2025) 100216

Contents lists available at ScienceDirect

Intelligence-Based Medicine
journal homepage: [Link]/journal/intelligence-based-medicine

Automatic glioma segmentation based on efficient U-net model using

MRI images
Yessine Amri a,b,* , Amine Ben Slama c , Zouhair Mbarki d , Ridha Selmi e , Hedi Trabelsi c
a
Biochemistry Laboratory, Bechir Hamza Children’s Hospital, Tunis, Tunisia
b
University of Jendouba, Higher Institute of Applied Studies in Humanity Le Kef, Department of Educational Sciences, Kef, Tunisia
c
Research Laboratory of Biophysics and Medical Technologies LR13ES07, Higher Institute of Medical Technologies of Tunis (ISTMT), University of Tunis EL Manar,
1006, Tunis, Tunisia
d
RIFTSI Research Laboratory, ENSIT, University of Tunis, 1008, Tunis, Tunisia
e
Northern Border University, College of Sciences, P. O. Box 1631, 91431, Arar, Saudi Arabia

A R T I C L E I N F O A B S T R A C T

Keywords: Gliomas are among the most aggressive and challenging brain tumors to diagnose and treat. Accurate segmen­
Glioma tation of glioma regions in Magnetic Resonance Imaging (MRI) is essential for early diagnosis and effective
MRI images treatment planning. This study proposes an optimized U-Net model tailored for glioma segmentation, addressing
Segmentation
key challenges such as boundary delineation, computational efficiency, and generalizability. The proposed model
Dice
integrates streamlined encoder-decoder pathways and optimized skip connections, achieving precise segmen­
U-net architecture
tation while reducing computational complexity. The model was validated on two datasets: TCGA-TCIA, con­
taining 110 patients, and the multi-modal BraTS 2021 dataset. Comparative evaluations were conducted against
state-of-the-art methods, including Attention U-Net, Trans-U-Net, DeepLabV3+, and 3D U-Net, using metrics
such as Dice Coefficient, Intersection over Union (IoU), Hausdorff Distance (HD), and Structural Similarity Index
(SSIM). The proposed U-Net achieved the highest performance across all metrics, with a Dice score of 92.54 %,
IoU of 90.42 %, HD of 4.12 mm, and SSIM of 0.962 on the TCGA-TCIA dataset. On the BraTS dataset, it achieved
comparable results, with a Dice score of 91.32 % and an IoU of 89.56 %. In contrast, other methods, such as
Attention U-Net and DeepLabV3+, showed lower Dice scores of 85.62 % and 84.10 %, respectively, and higher
HD values, indicating inferior boundary delineation. Additionally, the proposed model demonstrated compu­
tational efficiency, processing images in 1.5 s on average, compared to 5.0 s for Attention U-Net and 9.0 s for
Trans-U-Net. These results underscore the potential of the optimized U-Net as a robust, accurate, and efficient
tool for glioma segmentation. Future work will focus on clinical validation and extending the model to include
automated glioma grading, further enhancing its applicability in medical imaging workflows.

1. Introduction The accurate identification and delineation of gliomas are para­

mount for prognosis, treatment planning, and therapeutic evaluation.
Gliomas, a category of brain and central nervous system tumors, are Imaging modalities such as computed tomography (CT) and magnetic
among the most aggressive and life-threatening cancers. These tumors resonance imaging (MRI) have long been essential in diagnosing brain
account for a significant proportion of primary brain cancers, with tumors. MRI, in particular, is favored for its superior contrast resolution
survival rates heavily influenced by the stage at diagnosis and the pa­ and ability to non-invasively provide detailed information about tumor
tient’s age. In the United States, individuals diagnosed with brain or size, location, and structural abnormalities [2,3]. However, manual
central nervous system cancers face a five-year survival rate of only 36 segmentation of glioma regions by radiologists is inherently
%, underscoring the devastating nature of these malignancies. However, time-consuming, prone to inter-observer variability, and inadequate for
in younger populations, particularly children under the age of 15, the processing large datasets in clinical or research settings [4].
survival rate improves to approximately 75 %, reflecting the critical role In recent years, advancements in artificial intelligence (AI), partic­
of early detection and treatment [1]. ularly deep learning, have transformed medical image analysis.

* Corresponding author. Biochemistry Laboratory, BechirHamza Children’s Hospital, Bab Saadoun Square, 1007, Tunis, Tunisia.
E-mail address: [Link]@[Link] (Y. Amri).

[Link]
Received 13 November 2024; Received in revised form 18 January 2025; Accepted 26 January 2025
Available online 27 January 2025
2666-5212/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license ([Link]
nc-nd/4.0/).
Y. Amri et al. Intelligence-Based Medicine 11 (2025) 100216

Convolutional Neural Networks (CNNs) have proven highly effective in performance on BraTS dataset, demonstrating its robustness across
tasks ranging from image classification to segmentation, offering different glioma grades and imaging modalities.
unparalleled accuracy in identifying complex patterns within imaging
data [5]. Among CNN-based methods, Fully Convolutional Neural Net­ 2. Materials and methods
works (FCNNs) have demonstrated significant promise in biomedical
image segmentation. One of the most widely adopted architectures in 2.1. Data description
this domain is U-Net, which features a symmetric encoder-decoder
structure complemented by skip connections that preserve spatial res­ In this paper, the used dataset consists of a subset derived from The
olution [6–8]. Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA)
Building upon the U-Net model, researchers have explored various [15]. The data presents all modalities, resorting to FLAIR only when
enhancements to improve segmentation performance. For instance, other modalities were unavailable. It includes data from 110 patients
Kermi et al. [9] introduced refinements to the contracting and expand­ diagnosed with lower-grade glioma tumor (101 patients with all se­
ing paths of the U-Net architecture, demonstrating improved accuracy in quences available, and 9 patients with missing post-contrast sequence).
brain tumor segmentation tasks. However, challenges remain, particu­ The images were acquired with a resolution of 256 x 256 pixels, covering
larly regarding the reliance on large datasets for training and the axial slices of the brain with a slice thickness of 1 mm. In glioma cases,
computational inefficiencies of standard U-Net implementations, which contrast enhancement is particularly significant. Therefore, assessing
limit their feasibility for real-time clinical use. the tumor’s shape primarily relies on the abnormalities observed in
Subsequent variants, such as Attention U-Net, incorporated attention Fluid-Attenuated Inversion Recovery (FLAIR) imaging. The FLAIR
gates to focus on relevant features while suppressing irrelevant back­ technique suppresses fluid signals, allowing for clearer visualization of
ground noise [10,11]. Trans-U-Net combined Transformer layers with lesions near fluid-filled spaces (Fig. 1). The dataset was divided into 70
U-Net to capture long-range dependencies, enhancing performance on % for training (77 patients) and 30 % for testing (33 patients). Care was
complex segmentation tasks [12]. taken to ensure an even distribution of tumor grades and sizes across the
Other notable methods include DeepLabV3+ [13], which leverages subsets to prevent potential biases.
atrous convolutions for multi-scale feature extraction, and 3D U-Net
[14], designed for volumetric segmentation tasks using 3D convolutions. 2.2. Data preprocessing
Despite these advancements, challenges persist, such as handling class
imbalance, ensuring robustness to variations in image quality, and All images were preprocessed to ensure consistency, including
achieving real-time performance for clinical applications. However, resizing to 256 x 256 pixels, intensity normalization to a range of [0, 1],
these methods often involve increased computational complexity. and data augmentation techniques such as rotation, flipping, trans­
The objective of this study is to develop an optimized U-Net model lation, and zooming to address class imbalance and improve model
for glioma segmentation in MRI images, addressing specific challenges generalization.
such as boundary delineation, computational efficiency, and generaliz­ Given the variability in tumor sizes, oversampling was performed for
ability to multi-modal datasets. Unlike existing approaches that rely on underrepresented tumor categories to ensure a balanced dataset. Tumor
complex architectures, the proposed model integrates streamlined masks with smaller regions were augmented at higher rates to
encoder-decoder pathways, optimized skip connections, and parameter- compensate for their scarcity. These preprocessing steps were applied
efficient designs to achieve high segmentation accuracy with reduced uniformly to the training set, while the test set remained untouched to
computational cost. Furthermore, this study validates the model’s ensure unbiased evaluation.

Fig 1. Examples of the employed MRI dataset.

2
Y. Amri et al. Intelligence-Based Medicine 11 (2025) 100216

2.3. Computational environment and resources maps, retaining the most prominent features and allowing the
model to detect more information about the input image at
The experiments were performed using Python (version 3.12) as the different scales. This process is repeated over several levels,
primary programming language, with TensorFlow (version 2.16) and effectively transforming the input image into a dense representa­
Keras (version 2.16) frameworks for developing and training deep tion that encodes both local and global patterns relevant to glioma
learning models. Libraries such as NumPy, Pandas, Matplotlib, and regions.
Seaborn were employed for data manipulation, image processing, and 3. Decoder:
evaluation tasks. The experiments were run on a laptop equipped with • The decoder reconstructs the spatial dimensions by progressively
an Intel Core i7-1355U processor (up to 5.0 GHz, 12 MB cache), 16 GB upsampling the feature maps back to the original input size.
RAM, and an NVIDIA GeForce MX550 GPU with 2 GB of memory, • Deconvolution (or Transposed Convolution) Layers: These
operating on Windows 10. layers upsample the feature maps, reversing the effect of down­
sampling done by the encoder.
2.4. The developed U-net architecture • Up-Pooling Layers: Up-pooling layers restore the spatial resolu­
tion of the feature maps to match the input image size.
The Convolutional Neural Network (CNN) architecture is highly • The upsampled feature maps are combined with corresponding
effective for the involved task of identifying and segmenting regions feature maps from the encoder part via skip connections, which
affected by brain cancer [16,17]. In this study, we have employed the helps retain fine details from the original input image.
U-Net model, a variant of CNN, for the segmentation of glioma in MRI • Skip Connections: To preserve spatial context, feature maps from
images. The U-Net structure is particularly suitable for biomedical image the encoding path are concatenated with corresponding maps in
segmentation due to its encoder-decoder layout and the efficient use of the decoding path, facilitating accurate localization.
skip connections, which maintain critical spatial details throughout the 4. Output Segmented MRI Image: The final layer produces a binary
processing pipeline. image that highlights the segmented glioma ROI (Region of Interest).
This output mask highlights gliomas regions within the MRI images,
2.4.1. Mathematical framework as depicted in Fig. 2.
In the U-Net model, segmentation is framed as a mapping function,
fθ: I→S, where: 2.4.3. Training and hyperparameters
The Developed U-Net model was trained on The Cancer Genome
• I represents the set of input images within the domain Ω⊂Rd (with d Atlas (TCGA) dataset, with specific training parameters shown in
≥ 2 dimensions), Table 1. During training, several configurations were evaluated to
• S is the segmented output, derived as S ∈ I*, which contains the optimize segmentation performance, such as adjusting the initial
desired tumor regions in the observed image I0∈I. learning rate ε, batch size, and the dropout probability is set to 0.2. For
• fθ is a complex nonlinear function parameterized by θ, optimized the bias learned during training, it is part of the neuron’s pre-activation
during training to minimize a loss function and yield high-accuracy parameters.
segmentation. The Leaky ReLU (Rectified Linear Unit) function is a variant of the
standard ReLU activation function, and its equation is defined as:
In the training phase, the model’s parameters, encapsulated in θ, are
f(x) = max(αx, x) (1)
adjusted to minimize the error between the predicted segmentation and
the ground truth. The function fθ represents a series of transformations For each convolutional layer, the Leaky ReLU activation function was
achieved through convolutional layers, arranged to maximize feature applied with a constant α = 0.01 to control the negative slope of the
extraction and minimize computational expense. This multi-layer input x, aiding in convergence by maintaining the model’s ability to
structure allows high-level feature extraction without requiring fully learn complex features.
connected layers, thereby enhancing segmentation performance while
maintaining efficiency [18].
For further details on segmentation model principles, refer to pre­ 2.5. Evaluation metrics
vious research on segmentation architectures [19]. The CNN method has
been widely employed for brain cancer analysis with a focus on deep The performance of the proposed U-Net model was evaluated using
networks [20]. Pereira et al. [21] successfully identified brain cancer the following metrics:
regions using CNN on MRI datasets, leveraging convolutional kernels
and stacked convolutional layers. In this study, our U-Net approach, ⁃ Dice Coefficient: Measures the overlap between predicted and
optimized for glioma detection, effectively captures complex data pat­ ground truth segmentations, with higher values indicating better
terns from MRI images [22,23]. performance.
⁃ Intersection over Union (IoU): Assesses the ratio of the intersection to
2.4.2. U-net architecture design the union of predicted and ground truth segmentations.
The U-Net model architecture used for glioma segmentation follows ⁃ Hausdorff Distance (HD): Quantifies the maximum distance between
these main components: the predicted and ground truth boundaries, providing insight into
boundary precision.
1. Input MRI Image: The model takes an MRI image as input. ⁃ Structural Similarity Index (SSIM): Evaluates perceptual similarity
2. Encoder: between predicted and ground truth images by considering lumi­
• The encoder is responsible for detecting the context of the image nance, contrast, and structural information.
by down sampling it through a series of convolutional and max-
pooling layers. 2.6. Additional validation on BraTS dataset
• Convolutional Layers: Each convolutional layer applies a set of
filters to the input image, which helps in capturing spatial features To ensure the generalizability of the proposed U-Net model, we
such as boundary and textures. validated its performance on the BraTS 2021 dataset [24]. This dataset
• Max-Pooling Layers: After each convolutional block, a max- contains multi-modal MRI scans (T1, T1-contrast enhanced, T2, and
pooling operation reduces the spatial dimensions of the feature FLAIR) of lower-grade gliomas. For consistency with our primary

3
Y. Amri et al. Intelligence-Based Medicine 11 (2025) 100216

Fig 2. U-net Encoder-Decoder structure

highlight the relative advantages of our proposed model in segmenting

Table 1
glioma regions.
The hyper parameters of the used method.
In therapeutic applications, the Trans-U-Net [12] model has been
Initialization Leaky Dropout Training used for synapse multi-organs segmentation into 8 abdominal organ
ReLU
categories based on abdominal CT images; as shown in Fig. 4. Compared
Bias α p Initial Final ԑ Epochs Batch to different-organs segmentation, the pneumonia segmentation task in­
ԑ
volves predicting a binary mask instead of multiclass pixel-wise seg­
0.1 0.01 0.2 0.003 0.00003 100 128
mentation. The Trans-U-Net architecture is adapted by combining the
U-Net and Transformer systems.
dataset, only the FLAIR modality was used in this validation. The images
were resized to 256 x 256 pixels and preprocessed similarly to the TCGA 2.7.3. 3D U-net
dataset, including intensity normalization and data augmentation. The The 3D U-Net architecture extends the 2D U-Net design into three
validation subset consisted of 125 cases, and segmentation performance dimensions, making it ideal for processing volumetric medical imaging
was evaluated using Dice, IoU, Structural Similarity Index (SSIM), and data, such as MRI and CT scans. By using 3D convolutions and 3D
Hausdorff Distance (HD) metrics. pooling layers, the model effectively captures volumetric context, which
is critical for segmenting three-dimensional structures like brain tumors.
2.7. Comparative approach of automatic glioma ROI segmentation Skip connections between encoder and decoder layers ensure the pres­
ervation of spatial details, enabling precise boundary delineation.
The proposed model was evaluated against Attention U-Net, Trans- Despite its high computational cost, 3D U-Net remains a powerful tool
U-Net, DeepLabV3+, and 3D U-Net, well-known methods in medical for volumetric segmentation tasks, particularly in applications where
image segmentation. These models were implemented using TensorFlow the spatial relationship between slices is critical [14].
and trained under identical conditions to ensure a fair comparison.
Performance metrics, including Dice Coefficient, Intersection over 2.7.4. DeepLabV3+
Union (IoU), and Hausdorff Distance, and Structural Similarity Index DeepLabV3+ builds on the semantic segmentation capabilities of
(SSIM) were calculated for all models. DeepLabV3 by introducing an encoder-decoder architecture. This model
To validate the statistical significance of the performance differ­ combines atrous spatial pyramid pooling (ASPP) for multi-scale feature
ences, paired t-tests were conducted for Dice and IoU metrics, with a extraction with a decoder to refine the segmentation output. The ASPP
significance level of p < 0.05. module utilizes dilated convolutions with varying dilation rates,
enabling the model to capture features at multiple scales. The addition
2.7.1. Attention U-Net of a decoder improves the preservation of spatial details, particularly at
The Attention U-Net model [10], presented here as a comparison object boundaries, making DeepLabV3+ effective for segmenting com­
method, is designed to enhance segmentation accuracy by incorporating plex medical structures such as gliomas. Its modular design allows for
attention mechanisms. This model has shown promising results in flexibility in handling high-resolution images with minimal computa­
medical imaging and is used here to benchmark the performance of our tional overhead, positioning it as a benchmark for segmentation tasks in
proposed U-Net-based approach. medical imaging [13].
In abdominal CT images, the attention U-Net [11] was designed for
multi-class segmentation. Moreover, the system employed was adapted 3. Results and discussions
for pneumonia segmentation. Four down and up-sampling levels are
employed after the concatenation step. The ReLU activation function is In the present study, The proposed U-Net model achieved the highest
used in batch normalization. Additionally, additive attention and ReLU Dice Coefficient (92.54 %) and IoU (90.42 %) compared to Attention U-
attention activation are used. Indeed, stridden convolution layers and Net (85.62 % Dice, 82.53 % IoU), Trans-U-Net (81.24 % Dice, 77.39 %
bilinear interpolation were used for down-sampling and up-sampling, IoU), DeepLabV3+ (84.10 % Dice, 81.12 % IoU), and 3D U-Net (86.50 %
respectively (see Fig. 3). Dice, 84.30 % IoU) (Table 2). Hausdorff Distance analysis further
highlighted the model’s superior boundary delineation capabilities,
2.7.2. Trans-U-Net with the proposed U-Net achieving the lowest value of 4.12 mm. The
Similarly, the Trans-U-Net model, another comparison method, in­ proposed method showed also a SSIM score of 0.962 which demon­
tegrates Transformer layers with U-Net to leverage both spatial and strates its capability to produce segmentations that closely resemble
contextual information. This architecture is included for comparison to ground truth images in terms of structural and perceptual similarity.

4
Y. Amri et al. Intelligence-Based Medicine 11 (2025) 100216

Fig 3. Attention U-Net architecture.

Paired t-tests confirmed that the performance differences between the results are attained with respect to the accuracy of the segmentation
proposed U-Net and other models were statistically significant (p < stage using 30 % from the MRI dataset.
0.05). These results substantiate the robustness and accuracy of the The computation time of the Developed U-Net model was thoroughly
proposed model across diverse glioma segmentation tasks. The superior evaluated and compared with traditional Trans-U-Net and Attention-U-
performance is attributed to the streamlined encoder-decoder design, Net models. Results showed that the proposed U-Net model achieved
optimized skip connections, and reduced parameter complexity, which significantly better efficiency, processing each image in 1.5 s on average.
allow for efficient feature extraction and spatial detail preservation. This performance was notably faster than the comparative methods;
The IoU score further confirms this advantage by demonstrating the Trans-U-Net, Attention-U-Net, DeepLabV3+, and 3D U-Net, which
model’s ability to accurately segment glioma areas with minimal false required 5, 9, 6.5, and 8 s per image, respectively, due to their more
positives and false negatives. Compared to the comparative methods, complex architectures. The computational efficiency of the proposed U-
the modified U-Net architecture optimized parameters allowing efficient Net model was achieved through a combination of reduced parameter
feature extraction without the additional complexity of attention or count, streamlined encoder-decoder pathways, and optimized skip
transformer layers. Overall, these results underscore the effectiveness of connections. These enhancements reduced computational overhead
the developed U-Net model in delivering accurate and reliable seg­ while maintaining high segmentation accuracy, making the model more
mentation outcomes for glioma detection in MRI images. suitable for real-time applications. These findings indicate that our
Fig. 5 illustrates the original and segmented MRI images to present model achieves a balance between segmentation accuracy and pro­
the effectiveness of the employed approach. The current segmentation cessing efficiency, addressing a critical need in glioma ROI analysis.

5
Y. Amri et al. Intelligence-Based Medicine 11 (2025) 100216

Fig 4. Trans-U-Net architecture.

Net across different datasets and show its ability to perform well with
Table 2
multi-modal MRI data.
Performance metrics across models.
The preprocessing pipeline played a critical role in achieving the
Model Dice IoU Hausdorff Processing Structural reported performance of the proposed U-Net model. Normalization
Coefficient (%) Distance Time (s) Similarity
ensured consistent input data quality, allowing the model to focus on
(%) (mm) Index:
SSIM feature extraction rather than intensity variations. Data augmentation
improved the model’s ability to generalize by simulating real-world
Proposed U- 92.54 90.42 4.12 1.5 0.962
Net
variations, such as different tumor orientations, sizes, and anatomical
Attention U- 85.62 82.53 6.10 5.0 0.928 distortions. Addressing class imbalances by oversampling and aug­
Net menting smaller tumor regions contributed to the model’s accuracy in
Trans-U-Net 81.24 77.39 6.80 9.0 0.910 segmenting both small and large tumors effectively. These preprocessing
DeepLabV3þ 84.10 81.12 5.50 6.5 0.922
steps not only enhanced the model’s robustness but also improved its
3D U-Net 86.50 84.30 5.00 8.0 0.935
reproducibility, enabling its application to datasets with similar
characteristics.
The validation of the proposed U-Net model on the BraTS 2021 The U-Net architecture [25] has inspired numerous variants tailored
dataset yielded comparable results to those achieved on the TCGA for specific segmentation tasks, as demonstrated by recent advance­
dataset. The model achieved a Dice score of 91.87 % and an IoU of 89.94 ments. Azad et al. [26] provided a comprehensive review of U-Net’s
%, demonstrating high segmentation accuracy. The Hausdorff Distance success in medical imaging, emphasizing the impact of attention
of 4.12 mm indicates precise boundary delineation, and the average mechanisms and multi-scale designs. Wu et al. [27] proposed the
processing time of 1.6 s per image highlights the model’s computational UIU-Net, a nested U-Net model that excels in detecting small objects,
efficiency. These results are consistent with those obtained on the pri­ highlighting the importance of hierarchical feature extraction. Zhang
mary dataset, where the Dice score was 92.54 % and the IoU score was et al. [28] introduced BCU-Net, which integrates ConvNeXt and U-Net to
90.42 %. These results validate the generalizability of the proposed U- combine the benefits of convolutional and transformer-based

6
Y. Amri et al. Intelligence-Based Medicine 11 (2025) 100216

Fig 5. Visualization results of the Enhanced UNet and other network models in the glioma Dataset segmentation task.

architectures, enhancing segmentation accuracy. Zhao et al. [29] pre­ 4. Conclusion

sented NAU-Net, incorporating neighbor attention mechanisms to
improve feature selection, achieving remarkable results in diabetic The identification of the glioma region poses a crucial challenge in
retinopathy segmentation. These studies underscore the versatility of the assessment of tumor grade. The CNN methodology is noted for
U-Net and its derivatives, providing valuable insights that guided the achieving the detection of glioma ROI. Multiple methods deal with
optimization of our proposed model. problems related to the vanishing gradient during training. Therefore, in
Compared to recent U-Net advancements such as UIU-Net, BCU-Net, order to address these segmentation issues, this study employs the U-Net
and NAU-Net, our proposed model focuses on optimizing computational framework. The proposed segmentation approach demonstrates both
efficiency while maintaining high segmentation accuracy. By simpli­ the shortest computation time and superior dice and IoU values when
fying the encoder-decoder pathways and incorporating refined skip compared to current methods from the literature for gliomas segmen­
connections, our approach ensures robust performance without the tation using MRI images. However, there are several directions for
complexity of attention or nested mechanisms, making it suitable for future research to enhance the model’s applicability and robustness
clinical applications. further. First, we plan to extend the model to detect both low-grade and
The proposed U-Net model incorporates several architectural en­ high-grade gliomas by training and validating it on larger, multi-center
hancements tailored to address the unique challenges of glioma seg­ datasets. Incorporating a diverse range of tumor grades will improve the
mentation. One critical challenge is the accurate delineation of tumor model’s generalizability and ensure that it can handle various glioma
boundaries, which often exhibit irregular shapes and complex textures. subtypes effectively. Additionally, collaborations with radiologists and
To address this, the model utilizes optimized skip connections that clinicians will be established to evaluate the model’s performance in
enable the decoder to incorporate fine-grained spatial details from real-world scenarios. This assessment will focus on the model’s utility
earlier encoder layers. These connections ensure that critical boundary for diagnosis, treatment planning, and monitoring, ensuring its practical
information is preserved during the upsampling process, leading to more relevance and acceptance in clinical practice.
precise segmentation maps. Additionally, the encoder-decoder path­
ways were streamlined by reducing redundant layers, which improved CRediT authorship contribution statement
computational efficiency while retaining the model’s ability to capture
hierarchical features. This optimization also mitigates the risk of over­ Yessine Amri: Writing – original draft, Visualization, Methodology,
fitting, especially when working with relatively small datasets like ours. Investigation, Formal analysis, Conceptualization. Amine Ben Slama:
The use of Leaky ReLU activation (α = 0.01) further enhances gradient Writing – review & editing, Visualization, Methodology, Formal anal­
flow, reducing the likelihood of vanishing gradients during training. ysis, Conceptualization. Zouhair Mbarki: Writing – review & editing,
Collectively, these modifications enhance the model’s capacity to Visualization, Methodology, Formal analysis. Ridha Selmi: Writing –
handle the spatial variability of gliomas, improve segmentation accu­ review & editing, Visualization, Methodology. Hedi Trabelsi: Writing –
racy, and achieve faster processing times compared to other methods. review & editing, Supervision, Project administration,
The performance of the proposed U-Net model aligns with findings Conceptualization.
from recent studies that emphasize the importance of hybrid methods
and optimization techniques in brain tumor segmentation. Saifullah and Ethics approval and consent to participate
Dreżewski demonstrated significant improvements using Particle
Swarm Optimization-enhanced preprocessing (PSO) and ensemble This study was conducted following the principles of the Declaration
learning methods [30–32]. These approaches, combined with deep of Helsinki developed by the World Medical Association, as established
learning frameworks such as U-Net and DeepLabV3+, highlight the by the World Medical Association, and received approval from the
value of optimization strategies for segmentation tasks. Our model’s Human Ethics committee of Béchir Hamza Children’s Hospital of Tunis.
improvements, achieved through streamlined encoder-decoder path­ All procedures were carried out in compliance with relevant guide­
ways and optimized skip connections, complement these strategies by lines and regulations.
providing a lightweight and efficient alternative while maintaining high The requirement for informed consent was waived by Bechir Hamza
segmentation accuracy. Children’s Hospital Human Ethics committee.

7
Y. Amri et al. Intelligence-Based Medicine 11 (2025) 100216

Consent for publication [13] Khodadadi Shoushtari F, Sina S, Dehkordi ANV. Automatic segmentation of
glioblastoma multiform brain tumor in MRI images: using Deeplabv3+ with pre-
trained Resnet18 weights. Phys Med 2022;100:51–63. [Link]
Not Applicable. ejmp.2022.06.007.
[14] Liu X, et al. Development and validation of the 3D U-Net algorithm for
Availability of data and materials segmentation of pelvic lymph nodes on diffusion-weighted images. BMC Med Imag
2021;21(1). [Link]
[15] Buda M, Saha A, Mazurowski MA. Association of genomic subtypes of lower-grade
For Data evaluation please contact Dr. Ben Slama Amine at; [Link] gliomas with shape features automatically extracted by a deep learning algorithm.
nslama@[Link]. Comput Biol Med 2019;109:218–25. [Link]
compbiomed.2019.05.002.
[16] Walsh J, Othmani A, Jain M, Dev S. Using U-Net network for efficient brain tumor
Funding segmentation in MRI images. Healthcare Analytics 2022;2:100098. [Link]
org/10.1016/[Link].2022.100098.
[17] Khairandish MO, Sharma M, Jain V, Chatterjee JM, Jhanjhi NZ. A hybrid CNN-svm
No funding was received for this study. threshold segmentation approach for tumor detection and classification of MRI
brain images. IRBM Jun. 2021. [Link]
Declaration of competing interest [18] Liu X, Song L, Liu S, Zhang Y. A review of deep-learning-based medical image
segmentation methods. Sustainability 2021;13(3):1224. [Link]
su13031224.
The authors declare that they have no known competing financia­ [19] Allioui H, Mohammed MA, Benameur N, Al-Khateeb B, Abdulkareem KH, Garcia-
linterestsor personal relationships that could have appeared to influence Zapirain B, Maskeliūnas R. A multi-agent deep reinforcement learning approach for
enhancement of COVID-19 CT image segmentation. J Personalized Med 2022;12
the work reported in this paper.
(2):309. [Link]
[20] Slama AB, Sahli H, Amri Y, Trabelsi H. Res-Net-VGG19: improved tumor
Acknowledgment segmentation using MR images based on Res-Net architecture and efficient VGG
gliomas grading. Applications in Engineering Science 2023;16:100153. https://
[Link]/10.1016/[Link].2023.100153. 100153.
We are deeply grateful to all those who contributed to the success of [21] Pereira S, Pinto A, Alves V, Silva CA. Brain tumor segmentation using
this research project. convolutional neural networks in MRI images. IEEE Trans Med Imag 2016;35(5):
1240–51. [Link]
[22] Sahli H, Ben Slama A, Labidi S. U-Net: a valuable encoder-decoder architecture for
References liver tumors segmentation in CT images. J X Ray Sci Technol 2021:1–12. https://
[Link]/10.3233/xst-210993.
[1] Brancato V, Cerrone M, Lavitrano M, Salvatore M, Cavaliere C. A systematic review [23] Krithika Alias AnbuDevi M, Suganthi K. Review of semantic segmentation of
of the current status and quality of radiomics for glioma differential diagnosis. medical images using modified architectures of UNET. Diagnostics 2022;12(12):
Cancers 2022;14(11):2731. [Link] 3064. [Link]
[2] Śledzińska P, Bebyn M, Furtak J, Koper A, Koper K. Current and promising [24] Baid U, Ghodasara S, Mohan S, Bilello M, Calabrese E, Colak E, Farahani K,
treatment strategies in glioma. Rev Neurosci 2023;34(5):483–516. [Link] Kalpathy-Cramer J, Kitamura FC, Pati S, et al. The RSNA-ASNR-MICCAI BraTS
10.1515/revneuro-2022-0060. 2021 benchmark on brain tumor segmentation and radiogenomic classification.
[3] van Santwijk L, Kouwenberg V, Meijer F, Smits M, Henssen D. A systematic review arXiv preprint, arXiv:2107.02314 2021. [Link]
and meta-analysis on the differentiation of glioma grade and mutational status by arXiv.2107.02314.
use of perfusion-based magnetic resonance imaging. Insights into Imaging 2022;13 [25] Ronneberger Olaf, Fischer P, Brox T. U-net: convolutional networks for biomedical
(1):102. [Link] image segmentation. arXiv (Cornell University) 2015. [Link]
[4] van der Velden BHM, Kuijf HJ, Gilhuijs KGA, Viergever MA. Explainable artificial arxiv.1505.04597.
intelligence (XAI) in deep learning-based medical image analysis. Med Image Anal [26] Azad R, Aghdam EK, Rauland A, Jia Y, Avval AH, Bozorgpour A,
2022:102470. [Link] Karimijafarbigloo S, Cohen JP, Adeli E, Merhof D. Medical image segmentation
[5] Tran KA, Kondrashova O, Bradley A, Williams ED, Pearson JV, Waddell N. Deep review: the success of U-Net. IEEE Trans Pattern Anal Mach Intell 2024 Aug;21.
learning in cancer diagnosis, prognosis and treatment selection. Genome Med [Link]
2021;13(1). [Link] [27] Wu X, Hong D, Chanussot J. UIU-Net: U-Net in U-Net for infrared small object
[6] Fasterholdt I, et al. Value assessment of artificial intelligence in medical imaging: a detection. IEEE Trans Image Process 2022 Dec 15;32:364–76. [Link]
scoping review. BMC Med Imag 2022;22(1):187. [Link] 10.1109/TIP.2022.3221173.
022-00918-y. [28] Zhang H, Zhong X, Li G, Liu W, Liu J, Ji D, Li X, Wu J. BCU-Net: bridging ConvNeXt
[7] Fyllingen EH, Stensjøen AL, Berntsen EM, Solheim O, Reinertsen I. Glioblastoma and U-Net for medical image segmentation. Comput Biol Med 2023 Jun 1;159:
segmentation: comparison of three different software packages. PLoS One 2016;11 106960. [Link]
(10):e0164891. [Link] [29] Zhao T, Guan Y, Tu D, Yuan L, Lu G. Neighbored-attention U-Net (NAU-Net) for
[8] Kermi A, Mahmoudi I, Khadir MT. Deep convolutional neural networks using U-net diabetic retinopathy image segmentation. Front Med 2023 Dec 7;10:1309795.
for automatic brain tumor segmentation in multimodal MRI volumes. Brainlesion: [Link]
Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries 2019:37–48. [30] Saifullah S, Dreżewski R. Improved brain tumor segmentation using modified U-net
[Link] based on Particle Swarm optimization image enhancement. In: Genetic and
[9] Kermi A, Andjouh K, Zidane F. Fully automated brain tumour segmentation system evolutionary computation conference (GECCO ’24 companion); 2024. [Link]
in 3D-MRI using symmetry analysis of brain and level sets. IET Image Process 2018; org/10.1145/3638530.3654339.
12(11):1964–71. [Link] [31] Saifullah S, Dreżewski R. Automatic brain tumor segmentation using convolutional
[10] Das N, Das S. Attention-UNet architectures with pretrained backbones for multi- neural networks: U-net framework with PSO-tuned hyperparameters. Parallel
class cardiac MR image segmentation. Curr Probl Cardiol 2024;49(1):102129. Problem Solving from Nature - PPSN XVIII 2024;15150:333–51. Cham: Springer.
[Link] [32] Saifullah S, Dreżewski R. Brain tumor segmentation using ensemble CNN-transfer
[11] Wang S, Li L, Zhuang X. AttU-NET: attention U-net for brain tumor segmentation. learning models: DeepLabV3plus and ResNet50 approach. Lect Notes Comput Sci
In: Crimi A, Bakas S, editors. Brainlesion: glioma, multiple sclerosis, stroke and 2024;14835:340–54. [Link] Cham:
traumatic brain injuries. BrainLes 2021. Lecture notes in computer science, vol. Springer.
12963. Cham: Springer; 2022. [Link]
[12] Nouman M, Mabrok M, Rashed EA. Neuro-TransUNet: segmentation of stroke
lesion in MRI using transformers. arXiv preprint arXiv:2406.06017 2024. https://
[Link]/10.48550/arXiv.2406.06017.