Deep Learning-Based Medical Image Segmentation for Brain Tumors

Yudian Pan

University of Toronto, Trinity College, Toronto, Canada

Keywords: Deep Learning, Segmentation, Medical Image, U-Net.

Abstract: Brain tumors present a considerable health challenge, dramatically impacting both survival and quality of life.

This study introduces a improved deep learning approach for segmenting brain tumors in MRI scans,

intending to overcome the constraints of those existing approaches. The proposed model builds upon the

conventional U-Net architecture by incorporating the Convolutional Block Attention Module (CBAM),

designed to enhance the feature extraction capabilities. By integrating both channel-wise and spatial attention

mechanisms, this approach emphasizes relevant tumor regions while preserving structural detail. Experiments

evaluations on the TCGA Brain Tumor MRI dataset confirm the remarkable advantages of our UNet+CBAM

model compared to baseline approaches, achieving a Dice coefficient of 0.936 and an IoU of 0.882. This

proposed model successfully captures tumor boundaries with high precision and provides detailed

segmentation maps that could assist clinical diagnosis. While acknowledging the challenges posed by

computational complexity, this study makes a significant contribution to the advancement of automated brain

tumor segmentation technology, which holds considerable potential for practical applications in medical

settings. Subsequent studies will prioritize the optimization of the model for real-time applications and the

enhancement of its generalizability across a range of clinical settings.

1 INTRODUCTION

Brain tumors refer to the uncontrolled growth of cells

within brain tissue and its surrounding structures,

greatly impacting both survival rates and patients’

quality of life. Their incidence is on the rise globally,

especially in developed countries. Due to the

significant differences in biological characteristics,

clinical manifestations and treatment responses, early

and accurate diagnosis and precise treatment are

always a challenge. MRI has become a major tool for

brain tumor detection and evaluation due to its

advantages in soft tissue imaging, which can present

rich tumor information through different sequences,

such as T1, T2, FLAIR, etc., but the complexity and

heterogeneity of brain tumors in images still pose a

challenge in accurate segmentation and analysis.

segmentation and analysis difficulties (Litjens & van

Ginneken, 2017).

Traditional methods (e.g., threshold segmentation,

edge detection, region growing, etc.) are often

difficult to balance accuracy and stability when

dealing with brain tumors with different

https://orcid.org/0009-0007-9429-7165

morphologies; machine learning can improve some of

the performance, but it relies too much on hand-

designed features, making it difficult to fully explore

high-level information (Ronneberger & Brox, 2015).

Deep learning has achieved breakthroughs in the field

of medical image analysis by leveraging the

automatic feature extraction and pattern recognition

capabilities of multi-layer neural networks recently.

Among these, U-Net and its 3D variant have largely

improved segmentation accuracy and spatial detail

capture of biomedical images by virtue of the

advantages of multi-scale feature extraction, hopping

connectivity and 3D convolution (Çiçek &

Ronneberger, 2016). Related studies, such as the

work of DeepSeg and Amin et al (Amin & Hassan,

2023), have not only excelled in automated

segmentation of brain tumors, but also promoted the

application of multi-classification and segmentation

of MRI images (Zeineldin & Burgert, 2020).

This study aims to propose an efficient and

accurate brain tumor image segmentation approach

utilizing deep learning. Firstly, advantages and

disadvantages of existing mainstream models are

systematically sorted out and compared through

340

Pan, Y.

Deep Learning-Based Medical Image Segmentation for Brain Tumors.

DOI: 10.5220/0013689700004670

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

In Proceedings of the 2nd International Conference on Data Science and Engineering (ICDSE 2025), pages 340-347

ISBN: 978-989-758-765-8

literature research to provide a theoretical basis for

subsequent model design (Ronneberger & Brox,

2015). Secondly, by combining multi-scale feature

extraction with multi-modal information fusion, the

paper proposes an improved network structure and

training strategy to adapt to complex tumor

morphology and improve robustness, while exploring

lightweight and modular design to enhance real-time

performance (Gupta & Dayal, 2023). Again, data

augmentation, migration learning and semi-

supervised learningare helped address the issues

regarding limited labelled data, enhancing the

model’s generalization capacities (Litjens & van

Ginneken, 2017). Finally, large-scale experimental

validation is conducted on different datasets and

analysed in comparison with existing methods in

order to deeply explore its potential in clinical

applications (Isensee & Maier-Hein, 2019).

Key contributions of this research are mainly

manifested as follows:

1. Model Improvement: The number of layers and

parameters of the traditional U-Net architecture are

adjusted to enhance the recognition ability of

complex morphology of brain tumors by adding

convolutional and pooling layers.

2. Data experiments: Using the public brain tumor

MRI dataset, after data preprocessing and

enhancement, the improved U-Net is used for image

segmentation experiments.

3. Performance Comparison: The enhanced model

is compared to the original U-Net in brain tumor

segmentation, with evaluation metrics including

segmentation accuracy, Dice coefficient and

intersection ratio.

4. Discussion of results: experimental outcomes

are thoroughly analyzed, summarizing this model’s

advantages and shortcomings in terms of practical

applications, which provides a reference basis for

subsequent research.

2 RELATED WORKS

2.1 Overview of Traditional Methods

for Brain Tumor Segmentation

Early approaches to brain tumor segmentation over

the past few decades mostly depended on traditional

image processing techniques, including edge

detection, region growth, threshold segmentation, and

segmentation methods based on graph theory

. These

methods use manual features such as image grayscale

and texture to analyze the brain structure, which can

achieve a certain degree of segmentation effect in

simple cases, but when the tumor morphology is

complex or the tissue structure is more ambiguous, it

is often difficult for traditional methods to meet the

requirements of high precision. In addition,

traditional methods are more sensitive to the noise in

the image and are easily disturbed, resulting in

unstable segmentation results (Litjens & van

Ginneken, 2017). Therefore, in practical applications,

the robustness and automation level of these methods

need to be improved.

2.2 Discussion of Recent Advancements

Using Deep Learning

The advancement of deep learning technology have

triggered more researchers to begin to adopt

convolutional neural networks to address brain tumor

segmentation. Since U-Net was proposed, its

symmetric encoder-decoder structure and jump

connection design have greatly improved the

accuracy of image segmentation (Ronneberger &

Brox, 2015). To further process 3D medical images,

its variation has also been proposed and successfully

applied to volumetric data segmentation (Çiçek &

Ronneberger, 2016). Building on this, many scholars

have extensively explored the role of multi-scale

feature extraction along with attention mechanism

they have play in enhancing segmentation

performance. For example, the DeepSeg framework

developed by Zeineldin et al. achieved automatic

segmentation using FLAIR images (Zeineldin &

Burgert, 2020); Amin et al. and Gupta et al. put forth

a new network architecture for multiclass

segmentation of brain tumors (Amin & Hassan, 2023)

(Gupta & Dayal, 2023); meanwhile, Díaz-Pernas et al.

further optimized the segmentation accuracy using

multiscale convolutional neural networks. In addition,

Roy et al, Woo et al., and Fu et al. Improved the

network's ability to focus on key features by

introducing an attention module (Roy et al. 2018)

(Woo & Kweon, 2018) (Fu & Lu, 2019); while No

New-Net proposed by Isensee et al still maintains

high segmentation performance while simplifying the

network structure (Isensee & Maier-Hein, 2019).

These works show that deep learning-based methods

have achieved some huge progresses in the field of

brain tumor segmentation, demonstrating a strong

capacity to access fine lesion details.

2.3 Limitations of Existing Methods

However, despite their excellent performance in

segmentation accuracy, deep learning methods still

have some shortcomings. First, existing approaches

typically require an extensive amount of high-quality

Deep Learning-Based Medical Image Segmentation for Brain Tumors

341

labelled data, which is often scarce and expensive to

obtain in clinical practice (Litjens & van Ginneken,

2017). Second, some of the network structures are too

complex and computationally intensive, resulting in

long training time and high hardware requirements,

which is not conducive to generalization to practical

applications (Isensee & Maier-Hein, 2019). In

addition, although the attention mechanism can

improve the segmentation effect, it also increases the

parameters of the model and the training difficulty

(Roy et al. 2018; Woo & Kweon, 2018; Fu & Lu,

2019). There are also methods that are prone to miss

or missegmentation when dealing with edge details

and small-sized lesions (Zeineldin & Burgert, 2020;

Amin & Hassan, 2023; Gupta & Dayal, 2023).

Overall, how to improve this new model’s

lightweight and generalization capacities, and also

ensure the segmentation accuracy is still a difficult

problem to be solved.

2.4 Highlighting the Research Gap that

the Paper Addresses

This paper centers on improving current brain tumor

segmentation methods. Firstly, data enhancement

strategies (e.g. random horizontal flipping and

rotation) are used to extend the labelled data and

improve the model robustness (Feng & Wu, 2021).

Second, an improved U-Net network is designed to

combine multi-scale feature fusion and attention

mechanisms to enhance detail capture and maintain

lightweight (Ronneberger & Brox, 2015). The cross-

entropy loss function, Adam optimiser, and the

combination of pixel accuracy and IOU metrics are

used in training to evaluate the model, balancing

segmentation effect and computational efficiency

(Zeineldin & Burgert, 2020). In conclusion, this paper

fills the gap of brain tumor segmentation in terms of

data utilization, model design and application,

improves accuracy and efficiency, and is helpful for

clinical diagnosis (Isensee & Maier-Hein, 2019).

3 METHOD

3.1 Description of the Proposed Deep

Learning Approach

The method of deep learning proposed is mainly

developed based upon the improved U-Net

architecture, which realizes the fine segmentation of

medical images (e.g., brain MRI images) by

constructing a symmetric encoder-decoder network.

In the encoder part, a series of downsampling

modules consisting of convolution and ReLU

activation are utilized to gradually extract the multi-

scale features in the image, and at the same time, the

maximum pooling operation is used to reduce the

spatial size of the feature map; and in the decoder

section, the corresponding layers in the encoder are

fused with the up-sampled features in the decoder

through hopping connection, and then the image

resolution is recovered gradually by using the

transposed convolution, so as to retain more spatial

detail information. Additionally, to improve this new

model’s generalization ability and robust traits, data

enhancement strategies, for example, random level

flipping and random rotation, are introduced during

the training process, so as to expand the diversity of

training samples. Whole network is trained with the

cross-entropy loss function; besides the Adam

optimizer is used to continuously update the

parameters, which ultimately achieves high-precision

segmentation of the target region in medical images.

3.2 Network Architecture (U-Net)

The foundational UNet architecture is a widely

recognized and utilized model in the domain of

medical image segmentation. Introduced by

Ronneberger et al. (2015), UNet features a symmetric

encoder-decoder structure, which

comprises a

symmetric extending pathway for accurate

positioning and a contracted pathway for contextual

acquisition. The encoder path uses convolutional

layers with ReLU activation, followed by max-

pooling for downsampling, whereas the decoder path

recovers the spatial dimensions using convolutional

layers and upsampling, as seen in Table 1.

ICDSE 2025 - The International Conference on Data Science and Engineering

342

Table 1: The UNet model.

Layer Name Type Kernel

Size

Stride Padding Output

Channels

Activatio

Input - - - - 3 -

Downsample 1 Conv + ReLU + Conv + ReLU 3x3 1 1 64 ReLU

Downsam

le 2 Conv + ReLU + Conv + ReLU 3x3 1 1 128 ReLU

Downsam

le 3 Conv + ReLU + Conv + ReLU 3x3 1 1 256 ReLU

Downsam

le 4 Conv + ReLU + Conv + ReLU 3x3 1 1 512 ReLU

Downsample 5 Conv + ReLU + Conv + ReLU 3x3 1 1 1024 ReLU

Upsample Transpose Conv + ReLU 3x3 2 1 512 ReLU

Upsample 1 Conv + ReLU + Conv + ReLU 3x3 1 1 512 ReLU

sam

le 2 Conv + ReLU + Conv + ReLU 3x3 1 1 256 ReLU

sam

le 3 Conv + ReLU + Conv + ReLU 3x3 1 1 128 ReLU

Output Conv 1x1 1 0 2 Softmax

A sophisticated attention mechanism known the

Convolutional Block Attention Module (CBAM)

improves feature representation by using channel and

spatial attention in an orderly way. The Channel

Attention Module (CAM) and Spatial Attention

Module (SAM) are the two major parts of CBAM. As

illustrated in Figure 1, the spatial attention

mechanism works on “where” an essential feature is

situated, but the channel attention mechanism

concentrates on “what” is significant in a particular

feature map.

Figure 1: CAM framework (Picture credit: Original).

Mathematically, the Channel Attention Module

(CAM) can be described as formula (1):

𝑀





𝐹



=𝜎



𝑀𝐿𝑃

𝐴

𝑣𝑔𝑃𝑜𝑜𝑙



𝐹





𝑀𝐿𝑃𝑀𝑎𝑥𝑃𝑜𝑜𝑙



𝐹







(1)

The input feature map is represented by ( 𝐹), the

sigmoid function is shown by ( 𝜎), and the multi-

layer perceptron is referred to as MLP. The

operations for average-pooling and max-pooling are

denoted by AvgPool and MaxPool, respectively.

The Spatial Attention Module (SAM) can be

described as formula (2):

𝑀





𝐹



=𝜎



𝑓

×





𝐴

𝑣𝑔𝑃𝑜𝑜𝑙



𝐹



;𝑀𝑎𝑥𝑃𝑜𝑜𝑙



𝐹









(2)

Notably, (𝑓

×

) denotes a convolution operation

with a filter size of 7x7, and [;] here represents the

operation of concatenation.

Table 2: The structure of the CBAM module.

Layer Name Type Kernel

Size

Stride Padding Output Channels Activation

Channel

Attention

AdaptiveAvgPool + Conv +

ReLU + Conv

1x1 1 0 in_channels // ratio,

in_channels

Sigmoid

Spatial

Attention

Conv 7x7 or

3x3

1 3 or 1 2, 1 Sigmoid

The integration of the CBAM module into the

UNet architecture involves embedding the CBAM

module after each convolutional block in both the

downsampling and upsampling paths. Such an

integration allows the network to refine its focus on

the most informative features and regions at multiple

stages of the learning process. The CBAM module’s

structure is displayed in Table 2.

In the improved UNet+CBAM model, each

downsampling block involves convolutional layers

followed by a CBAM module, and similarly, each

upsampling block includes convolutional layers

followed by a CBAM module. Through this

connection, this model’s capacity to recognise

intricate correlations and characteristics in the

medical images becomes stronger, and this leads to an

increase in the precision of segmentation, as shown in

the data below in Table 3.

Deep Learning-Based Medical Image Segmentation for Brain Tumors

343

Table 3: This caption needs to be modified to justify as it includes many lines.

Layer Name Type Kernel

Size

Stride Padding Output

Channels

Activation

Input - - - - 3 -

Downsample 1 Conv + ReLU + Conv + ReLU + CBAM 3x3 1 1 64 ReLU

Downsample 2 Conv + ReLU + Conv + ReLU + CBAM 3x3 1 1 128 ReLU

Downsample 3 Conv + ReLU + Conv + ReLU + CBAM 3x3 1 1 256 ReLU

Downsample 4 Conv + ReLU + Conv + ReLU + CBAM 3x3 1 1 512 ReLU

Downsample 5 Conv + ReLU + Conv + ReLU + CBAM 3x3 1 1 1024 ReLU

Upsample Transpose Conv + ReLU 3x3 2 1 512 ReLU

Upsample 1 Conv + ReLU + Conv + ReLU + CBAM 3x3 1 1 512 ReLU

Upsample 2 Conv + ReLU + Conv + ReLU + CBAM 3x3 1 1 256 ReLU

Upsample 3 Conv + ReLU + Conv + ReLU + CBAM 3x3 1 1 128 ReLU

Output Conv 1x1 1 0 2 Softmax

3.3 Preprocessing Techniques

Before feeding the images into the network, this work

employs multiple types of methods for preprocessing

steps to standardize the input data. Images are first

resized to a fixed dimension (e.g., 256×256). In

addition, intensity normalization is applied to adjust

the pixel values to contain zero mean and unit

variance. The normalization formula is as formula (3):

𝑥=

𝑥−𝜇

𝜎

(3)

In this equation, 𝑥 is the original pixel value, 𝜇 is

the mean, and 𝜎 is the standard deviation computed

over all pixels of the image. This normalization step

improves the convergence of the network during

training.

segmentation masks. The dataset is denoted by

formula (4):

𝐷=



𝑥



,𝑦









(4)

Where 𝑥



represents the 𝑖 th input image; 𝑦



its

related ground truth mask, with N being the total

number of samples. The training process is guided by

the cross-entropy loss function, being defined as

formula (5):

𝐿=−

𝑁





















𝑦









𝑙𝑜𝑔 𝑙𝑜𝑔 𝑝











(5)

Notably, 𝐶denotes the number of classes, 𝑦









the true label for class 𝑐 (usually 0 or 1), and 𝑝









the predicted probability for that class. The paper

optimizes the network parameters by means of the

Adam optimizer with a learning rate typically set

to 1×10



. This combination of loss function and

optimizer has proven effective in training deep

segmentation models.

3.4 Dataset Description

The TCGA Brain Tumor MRI dataset is derived from

the TCGA (The Cancer Genome Atlas) project.

Which collects brain MRI images from real clinics

and provides the corresponding tumor segmentation.

Due to the wide range of data sources, the samples

cover different types, sizes and morphologies of brain

tumors, making this dataset an important benchmark

data in the segmentation of medical images. The

diversity of the data also contributes to this model’s

increased resilience and generalization performance.

Researchers can use this dataset to train and verify

deep learning models for better automatic

identification and fine segmentation of brain tumor

lesions.

3.5 Loss Function, Optimizer, and

Learning Rate

In this study, the paper uses the cross-entropy loss

function as the optimization objective for the

segmentation task. The cross-entropy loss measures

the discrepancy between the predicted probability

distribution and the true label distribution. Its formula

is given by equation (6):

𝐿=−

𝑁





















𝑦









𝑙𝑜𝑔 𝑙𝑜𝑔 𝑝











(6)

Here, N denotes the total number of samples, C

represents the number of classes, 𝑦









is the true label

ICDSE 2025 - The International Conference on Data Science and Engineering

344

(either 0 or 1) for class c in the ith sample, and 𝑝









the probability predicted by the model for that class.

To efficiently update the model parameters, this paper

resorts to the Adam optimizer, which adaptively

adjusts learning rate during training. This paper’s

experiments reveal that the initial learning rate is

designed to 𝑏𝑒 1 × 10



and is fine-tuned based on

validation results to balance training stability and

model performance.

3.6 Evaluation Metrics for

Segmentation Performance

The Dice coefficient is used through the paper to

assess our segmentation model’s performance. The

area that overlaps between the ground truth and the

predicted segmentation zone is measured by this

statistic. Formulas (7) and (8) are applied to calculate

the dice coefficient:

𝐼𝑜𝑈 =

𝑃∩𝐺

𝑃∪𝐺

(7)

𝐷𝑖𝑐𝑒 =

2×

𝑃∩𝐺

𝑃



𝐺

(8)

The number of pixels in each set is indicated by the

formula, where P stands for the set of pixels in the

predicted segmentation and G for the set of pixels in

the ground truth mask. A greater match between the

forecast and the ground truth is demonstrated by a

larger value of the Dice coefficient, which spans from

0 to 1.

4 RESULTS

4.1 Performance Metrics Compared

with State-of-the-art Methods

Comparative tests on several popular models for the

brain tumor segmentation had been carried out in the

paper. The experiments compared the traditional

UNet base model, DeepLabV3, PSPNet, and MA-Net,

and on the basis of which the CBAM module was

embedded into UNet to form the UNet+CBAM (Full)

model. According to the experimental findings, the

IoU is 0.854 and the Dice coefficient of the UNet base

model is 0.915, respectively; the Dice coefficient of

DeepLabV3 is 0.924, and the IoU is 0.866; the Dice

coefficient of PSPNet is 0.920, and the IoU is 0.861;

the Dice coefficient of MA-Net is 0.928, and the IoU

is 0.873; while our proposed UNet+CBAM (Full)

model achieves a Dice coefficient of 0.936 and an

IoU of 0.882. From those data, it is evident that the

model’s segmentation effect has greatly enhanced

when the CBAM module is integrated. Table 4 below

summarizes the findings of the comparison

experiments:

Table 4: Main results.

Model Dice Score IoU

UNet Base 0.915 0.854

Dee

LabV3 0.924 0.866

PSPNet 0.920 0.861

MA-Net 0.928 0.873

UNet+CBAM (Full) 0.936 0.882

4.2 Visualizations of Segmentation

Outputs

The segmentation findings of the final model are

illustrated in this paper. The model can precisely

identify the region of the brain tumor with distinct

segmentation borders and detailed features, as shown

in Figure 2 below. Both the overall contour of the

tumor and the internal structural details are well

captured by the model and restored on the output

image. These visualization results fully demonstrate

the good performance of the model in the actual

segmentation task, as well as providing intuitive

support for subsequent clinical aid diagnosis, as

shown in Figure 2.

Figure 2: Main result (Picture credit: Original).

Deep Learning-Based Medical Image Segmentation for Brain Tumors

345

4.3 Analysis and Interpretation of

Results

From the indicators of the segmentation results, the

final model shows high accuracy and robustness. The

model’s ability to precisely identify and segment the

tumor region under a range of conditions is

demonstrated by the high degree of overlap between

the actual annotations and predicted findings shown

in the images. By observing the segmentation results,

the paper find that the model is particularly

meticulous in processing the edges of the tumor

region, and is able to effectively capture the subtle

changes of the lesion. This fully demonstrates that the

designed network structure and attention module play

an active role in feature extraction and information

fusion, making this model highly effective and

reliable in segmentation in practical applications.

4.4 Discussion of Potential limitations

Although the final model performs exceptionally well

in the segmentation task, there might still be

limitations. First, although the dataset used is highly

representative, more cases with different scanning

devices, different imaging parameters and variable

image quality may be encountered in practical

applications, which may have some impact on the

model performance. Secondly, the computational

complexity of the model is high, and although it runs

smoothly in the experimental environment, the real-

time and hardware requirements still need to be

further considered in practical clinical applications.

The adaptability and operational efficiency of the

model can be further improved in future work through

model pruning, lightweight design, and more data

enhancement means.

5 CONCLUSIONS

A brain tumor segmentation method based on

improved U-Net and attention mechanism is

proposed in this paper. Experiments show that the

new model has significant improvement in both Dice

coefficient and IoU, and the segmented tumor region

has clear edges and rich details. The method not only

reduces the workload of manual labelling, but also

provides an intuitive segmentation reference for

doctors, which has certain clinical application

potential.

Meanwhile, the study also exposed some

shortcomings. The current dataset has limited sample

sources and the model is computationally large,

which may require further optimisation in real

scenarios. In the future, the paper can try to introduce

more data, adopt semi-supervised or migration

learning methods, or design a more lightweight

network structure to boost the stability and operation

efficiency of this model.

Overall, this paper provides a new idea for

automatic brain tumor segmentation, and the paper

expect that it will play a greater role in clinical

practical applications in the future.

REFERENCES

Amin, B., Samir, R. S., Tarek, Y., Ahmed, M., Ibrahim, R.,

Ahmed, M., & Hassan, M. 2023. Brain tumor multi-

classification and segmentation in MRI images using

deep learning. arXiv preprint arXiv:2304.10039.

Çiçek, Ö., Abdulkadir, A., Lienkamp, S. S., Brox, T., &

Ronneberger, O. 2016. 3D U-Net: Learning dense

volumetric segmentation from sparse annotation. In

International Conference on Medical Image Computing

and Computer-Assisted Intervention (424–432).

Springer, Cham.

Díaz-Pernas, F. J., Martínez-Zarzuela, M., Antón-

Rodríguez, M., & González-Ortega, D. 2024. A deep

learning approach for brain tumor classification and

segmentation using a multiscale convolutional neural

network. arXiv preprint arXiv:2402.05975.

Feng, R., Liu, X., Chen, J., Chen, D. Z., Gao, H., & Wu, J.

2021. A deep learning approach for colonoscopy

pathology WSI analysis: Accurate segmentation and

classification. IEEE Journal of Biomedical and Health

Informatics, 25(10), 3700–3708.

Fu, J., Liu, J., Tian, H., Li, Y., Bao, Y., Fang, Z., & Lu, H.

2019. Dual attention network for scene segmentation.

In Proceedings of the IEEE/CVF Conference on

Computer Vision and Pattern Recognition (3146-3154).

Gupta, A., Dixit, M., Mishra, V. K., Singh, A., & Dayal, A.

2023. Brain tumor segmentation from MRI images

using deep learning techniques. arXiv preprint

arXiv:2305.00257.

Isensee, F., Kickingereder, P., Wick, W., Bendszus, M., &

Maier-Hein, K. H. 2019. No new-net. In International

MICCAI Brainlesion Workshop (234–244). Springer,

Cham.

Kamnitsas, K., Ledig, C., Newcombe, V. F., Simpson, J. P.,

Kane, A. D., Menon, D. K., … & Glocker, B. 2017.

Efficient multi-scale 3D CNN with fully connected

CRF for accurate brain lesion segmentation. Medical

Image Analysis, 36, 61–78.

Litjens, G., Kooi, T., Bejnordi, B. E., Setio, A. A. A.,

Ciompi, F., Ghafoorian, M., … & van Ginneken, B.

2017. A survey on deep learning in medical image

analysis. Medical Image Analysis, 42, 60–88.

ICDSE 2025 - The International Conference on Data Science and Engineering

346

Ronneberger, O., Fischer, P., & Brox, T. 2015. U-Net:

Convolutional networks for biomedical image

segmentation. arXiv preprint arXiv:1505.04597.

Roy, A.G., Navab, N., Navab, N., et al. 2018. Concurrent

spatial and channel squeeze and excitation in fully

convolutional networks. In: Proceedings of the medical

image computing and computer-assisted intervention,

Singapore, MICCAI 2018, 421–429.

Woo, S., Park, J., Lee, J.-Y., & Kweon, I. S. 2018. CBAM:

Convolutional Block Attention Module. In:

Proceedings of the European Conference on Computer

Vision (ECCV).

Zeineldin, R. A., Karar, M. E., Coburger, J., Wirtz, C. R.,

& Burgert, O. 2020. DeepSeg: Deep neural network

framework for automatic brain tumor segmentation

using magnetic resonance FLAIR images. arXiv

preprint arXiv:2004.12333.

Deep Learning-Based Medical Image Segmentation for Brain Tumors

347