Enhanced YOLOv8 Framework for Early Detection of

Alzheimer's Disease Using MRI Scans

Safa Jraba

, Mohamed Elleuch

, Hela Ltifi

and Monji Kherallah

National School of Electronics and Telecommunications (ENETCom), University of Sfax, Tunisia

National School of Computer Science (ENSI), University of Manouba, Tunisia

Faculty of Sciences and Techniques of Sidi Bouzid, University of Kairouan, Tunisia

Faculty of Sciences, University of Sfax, Tunisia

Keywords: YOLO, Deep Learning, Early Diagnosis, MRI, Medical Imaging, Detection, Brain Imaging, YOLOv8.

Abstract: Alzheimer's disease is characterized by a progressive neurodegenerative disorder, often misdiagnosed too

late, with early symptoms that are hidden. Detection is crucial for effective treatment and slowing the

progression of disease. We propose an upgraded version of the YOLO (You Only Look Once) framework,

namely YOLOv8, for detecting Alzheimer's disease from MRI scans. Our approach seeks the detection of

early structural changes in the brain, most particularly in the hippocampus and cortex, which are also among

the first areas affected in this disease process. The framework performs state-of-the-art detection of

Alzheimer's changes with a 96% precision via multi-scale feature extraction specifically designed for

neuroimaging data. Results show this approach to be exceptionally effective in improving sensitivity and

precision over existing techniques, marking it as a highly reliable method for early diagnosis of Alzheimer's

disease.

1 INTRODUCTION

Alzheimer's disease (AD) is reportedly the most

common form of dementia, affecting millions across

the world. In order to control the disease, it is most

essential to diagnose it in an early stage for timely

intervention; however, subtle structural changes

within the brain caused by the disease, especially in

the early stages of Alzheimer's, are often missed by

traditional diagnostic techniques. This study

showcases an enhanced YOLO-based framework

developed for the early diagnosis of Alzheimer's

using MRI data with special emphasis on structural

brain abnormalities, with a view to providing

effective measures of hippocampal atrophy and

cortical thinning.

The automatic identification of brain tumors from

Magnetic Resonance Imaging (MRI) is a highly

challenging and labor-intensive task. Amine et al.

(2022).

https://orcid.org/0009-0007-7818-5091

https://orcid.org/0000-0003-4702-7692

https://orcid.org/0000-0003-3953-1135

https://orcid.org/0000-0002-4549-1005

Prompt detection of brain tumors is very

important for successful treatment outcomes and

better prognoses of the patients. Consequently, the

identification of brain tumors plays a vital role in

medical diagnostics. Magnetic Resonance Imaging

(MRI) is regarded as the best imaging technique for

visualizing the brain and detecting the tumors. The

You Only Look Once (YOLO) series have witnessed

promising results in accurately detecting brain

tumors. For instance. Kang et al. (2023), proposed

RCS-YOLO, a novel YOLO framework

incorporating reparameterized convolution with

channel shuffle specially proposed for brain tumor

detection, achieving a good trade-off between speed

and accuracy.

The modules Conv. King et al. (2023). C2f

(shortcut), and Spatial Pyramid Pooling Fast (SPPF)

make up the backbone, which is in charge of feature

extraction. The Conv and SPPF are comparable to

those found in the YOLOv5 architecture. Jocher et al.

Jraba, S., Elleuch, M., Ltiﬁ, H. and Kherallah, M.

Enhanced YOLOv8 Framework for Early Detection of Alzheimer’s Disease Using MRI Scans.

DOI: 10.5220/0013315300003890

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

In Proceedings of the 17th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2025) - Volume 3, pages 1229-1237

ISBN: 978-989-758-737-5; ISSN: 2184-433X

1229

(2022). Conv is also known as ConvBiSiLU (or

CBS). Alongside C2f (shortcut), the Conv module

performs convolution operations on input pictures to

facilitate feature extraction, while SPPF permits an

adjustable output size. The C2f (shortcut)

convolutional structure is lighter than the C3 module

of YOLOv5. The head and the backbone are the two

main components of the YOLOv8 architecture Jocher

et al. (2023), with the neck being incorporated into

the head part. King et al. (2023).

Researchers and medical professionals can locate

the area of the brain afflicted by a tumor by using

MRI, an imaging method that shows the anatomy and

structure of the human brain Sakthidasan et al.

(2021).

2 RELATED WORKS

The search for precise detection methodologies for

cerebral tumors has sparked significant research

efforts in recent years. Numerous approaches have

been investigated to address this vital need in the field

of health. Conventional diagnostic modalities such as

magnetic resonance imaging (MRI) and computed

tomodensitometry (CT scans) have historically been

the primary tools used to identify brain tumors. Their

effectiveness in early detection and precise

delineation of terrorist borders, however, remains a

challenge.

The precise Accurate automated classification of

brain MRI images is crucial in medical research,

distinguishing healthy brain tissue from tumor-

affected areas (benign or malignant), as noted by

Gurbină and al. (2019).

CNN applied to MRI images has been shown to

be beneficial in many recent studies for the

classification of brain-related illnesses. Yuan et al.

(2018)

As such, these tasks are very prone to some

missed, misinterpreted, or wrong tumor-like

structures. Research currently aims mostly at the

classification and segmentation of tumors in MRI

scans, while the detection of tumors is relatively

underexplored. Although different CNNs have shown

very promising results in the field of brain tumor

detection, articles on the performance of You Only

Look Once (YOLO) networks in this regard are not

frequently published. Nevertheless, the architectures

of such networks are continuously becoming more

complex. Lather et al. (2020)

A thorough Waquas et al. (2020) provide a

comprehensive review of deep learning models for

brain tumor analysis, detailing datasets,

methodologies, evaluation standards, and current

detection approaches.

Amarapur et al. (2019) explored traditional

machine learning and deep learning methods for brain

tumor validation, using three algorithms to achieve

effective classification with improved accuracy and

robustness.

A high-level System Cancer Diagnosis by

coalescing the four Level-I taxonomy components

"DIV" Data, Image Segmentation processing and

VIEW is proposed in Lukampe al. (2019).

A novel deep learning framework driven by the

internet of health things (IoHT) for brain tumor

detection and tumor cell classification was presented

by the authors in Devunooru and al. (2021). Using the

common Pap smear Herlev dataset, the conventional

Machine Learning (ML) techniques—KNN, RF, NB,

LR, and SVM classifiers—are applied.

Archana et al. (2023) compared CNN optimizers

for brain tumor detection, finding AlexNet with

Adam achieved 94.76% accuracy on a dataset of 1547

images, outperforming LeNet with SGD.

Wani et al. (2023) explored brain tumor diagnosis

using MRI with CNNs like AlexNet, GoogleNet,

VGG-19, a bespoke model, and machine learning

models. Their hybrid approach achieved 90.625%

accuracy, highlighting the potential of combining

deep learning and traditional methods.

Rajinikanth et al. (2022) used pooling techniques

with pre-trained VGG16 and VGG19 CNNs to

classify glioma and glioblastoma from MRI images,

achieving 96.08% accuracy with average pooling and

VGG16. Modern developments, including YOLO

models for efficient object detection, will be reviewed

and compared to CNN performance.

Li and al. (2021) created the first YOLO as a real-

time object detector. The capacity to identify objects

with a single pass is the primary feature; each cell in

the grid-based image predicts possible object classes

and boxes. For real-time applications, Li’s initial

iteration of YOLO functioned as an object detector.

Kumar et al. (2020) presented YOLOv4 for

detecting brain tumors on a private local MRI

database. They achieved 90% accuracy and 88%

recall by employing several data augmentation

techniques and fine-tuning the YOLOv4 model.

Using YOLOv5 on Brain Anomalies: Chen et al.

(2022), on a private MRI dataset, used YOLOv5 for

the detection of small brain anomalies.Their model

achieved 92% accuracy and 90% recall, indicating

that YOLOv5 is most suitable for applications in

medicine, requiring highly precise detection.

An enhanced version of YOLOv5 for early

detection of AD was attempted by Park et al. (2023)

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1230

using BRATS 2020 data. In this study, the model

modified to accommodate accurate segmentation

yielded 93% accuracy and an F1 score of 0.92 for

brain-region detection in AD.

The YOLOv7 model was trained by Singh et al.

(2023) on the AMNI and BRATS 2019 datasets-they

employed transfer learning to make use of large

medical imaging databases. The model yielded an

accuracy and recall of 94% and 93% respectively.

Consequently, Zhao et al. (2021) investigated the

use of ResNet for bilateral brain abnormality

identification using the BRATS database, and they

were able to detect tumors with an MRI image

classification accuracy of 87%.

As far as themselves, that created opportunities

for Zhao et al. (2021) research into using ResNet to

identify bilateral brain abnormalities with the use of

the BRATS database. Using MRI image

classification, 87% classification accuracy was

attained for tumor recognition.

In this section, we look closely at a few recently

published, more successful techniques (See Table 1).

Table 1: Research on the detection of brain tumors with

YOLO.

Authors YOLO Model Dataset Acc.

Li et al.

(2021)

YOLOv3 with

attention layers for

important regions

BRATS 2018 88%

Kumar

et al.

(2020)

YOLOv4

University Hospital

(local data)

YOLOv4

optimized by

MRI data

augmentation

90%

Chen et

al.

(2022)

YOLOv5 with

multiscale

detection for small

tumors

Private MRI

Dataset

92%

Park et

al.

(2023)

Enhanced

YOLOv5

with segmentation

and detection

tuning

BRATS 2020 93%

Singh et

al.

(2023)

YOLOv7 with

transfer learning

and normalization

ADNI &

BRATS 2019

94%

Zhao et

al.

(2021)

YOLOv3-tiny for

optimized real-time

detection

Open Access

Dataset

80%

A summary of recent research on brain tumor

detection utilizing several YOLO models may be

found in Table 1. This table shows how the YOLO

architecture has been improved and refined for the

purpose of detecting brain tumors in MRI scans, as

well as how accuracy has increased across various

YOLO iterations and configurations.

3 METHODOLOGIES

This section provides an overview of the

experimental flow for developing Alzheimer's

disease-motivated experiments.

3.1 Gathering Data

The Alzheimer Disease Neuroimaging Initiative is an

internationally recognized Alzheimer's disease

research database.

The characteristics of the YOLOv8 model

facilitate easy training, validation, and evaluation

because there are three separate datasets for training,

validation, and testing.

Folder Structure: The dataset includes three main

folders: train, valid, and test, each containing

subfolders for images and labels, formatted for

YOLOv8 compatibility in real-time object recognition.

MRI Images: The dataset includes MRI slices of

brain regions, aiding in visual recognition of tumor-

related characteristics and abnormalities. It features

an example image highlighting colorful brain regions

with distinct textures indicative of potential

pathologies.

Annotations: Each image is paired with a .txt file

containing annotations in YOLOv8 format, indicating

the class (0 for "normal," 1 for "tumor," 2 for other

anomalies) and the bounding box's normalized

coordinates and dimensions (0 to 1), centered on the

areas of interest.

There are three primary classes defined:

 Images: that provide no outward indications of

abnormalities are considered normal (class 0).

 Tumor (Class 1): Pictures that clearly display

brain tumors.

 Possible Anomaly (Class 2): Pictures with

obvious symptoms that call for medical attention

if an abnormality diagnosis is necessary.

Table 2: MRI Image categorization for the Detection of

Brain Variations.

Class

Name

Description

0 Normal

Images with no visible indications

of abnormalities.

1 Tumor

Images that clearly display brain

tumors.

Possible

Anomaly

Images showing symptoms that

suggest a possible abnormality,

uirin

medical attention.

Enhanced YOLOv8 Framework for Early Detection of Alzheimer’s Disease Using MRI Scans

1231

The three primary classifications established for

the classification of brain MRI images are

summarized in this table.

3.2 Pre-Processing

Our brain tumor MRI dataset underwent a number of

pre-processing procedures to guarantee data

consistency and quality. In order to fully utilize the

features in the image, the data was kept as raw as

possible and no augmentation was made. The image's

640x640 pixel size makes sense considering the

trade-off between detection precision and time-

constraint compliance.

Figure 1: Sample MRI Images from Brain Tumor Detection

Dataset.

The six sample MRI images displayed in figure 1

are from the dataset used for brain tumor detection.

These images illustrate the diversity of the brain

tumor cases to be analyzed by the YOLOv8

framework. The dataset provides the basis for training

and testing the detector on variations in structure and

potential tumor locations when identifying and

localizing brain tumors.

3.3 Proposed Architectures

3.3.1 YOLO v8 Architecture

The YOLO series, a well-liked one-stage detection

algorithm, is very good at object detection tasks

because it strikes a compromise between speed and

accuracy. A C2f is adopted by YOLOV8.

It streamlines the procedure and speeds up

detection by separating the classification and

detection heads using an Anchor Free head. Luo et al.

(2023).

Figure 2 describes the framework of the YOLOv8

model in its adaption for Alzheimer's disease

detection altered by MRI scans. Three main

categories: Backbone, Neck, and Output can be used

to categorize the model's structure.

Figure 2: YOLO v8 architecture.

3.3.2 Backbone

The features of the input MRI image are extracted by

the backbone. It begins with convolutional layers

(Conv) and moves on to more complex layers like C2f

and Fast_C2f layers, which aim to improve the

feature extraction process by preserving the spatial

information crucial for identifying Alzheimer's

markers. An input image has been resized 640x640

pixels along with the three colors channels

(640x640x3) so that all of the model's architecture

can be executed uniformly.

Elharrouss et al. (2022) examines diverse

backbone architectures and their evolution, as well as

their applicability in the extraction of features for

intricate deep learning problems.

3.3.3 Neck

Features taken from various Backbone levels are

further refined and combined by the AFPN

(Augmented Feature Pyramid Network) layer, which

is located in the Neck.

Al-Nawashi et al. (2023) examines how the

addition of multi-scale feature fusion modules,

similar to AFPN, can improve contextual

understanding of spatial relationships in MRI images,

enhancing the model's ability to detect subtle signs of

Alzheimer's disease.

3.3.4 Output

To enable multi-scale detection, the output layer is

made up of three detection heads (P3, P4, and P5) that

are stacked to feature maps of varying sizes (80x80,

40x40, and 20x20). Chen et al. (2024).

Two modules further process them: Cls

classification based on BCE (Binary Cross-Entropy)

loss, which maximizes accuracy in detecting

Alzheimer's disease, and Bbox for bounding box

predictions using metrics like CIoU (Complete

Intersection over Union) and DFL (Distribution Focal

Loss) for spatial accuracy. Luo et al. (2024).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1232

4 EXPERIMENTS AND RESULTS

In the present study, we present an advanced model

based on the YOLOv8 architecture for the early

detection of Alzheimer's disease using MRI images.

The experimental protocol is organized to evaluate

the ability of the YOLOv8 model to accurately detect

Alzheimer's disease markers in MRI images.

This network architecture has been trained and

tested on a dataset of MRI scans, resized to 640x640

pixels over three color channels. The YOLOv8

backbone efficiently extracts essential features from

MRI images, while its multi-scale feature detection

capabilities guarantee robust identification of subtle

and pronounced markers.

The enhanced system integrates AFPN

(Augmented Feature Pyramid Network) for enhanced

contextual awareness, and uses detection heads (P3,

P4 and P5) to predict boundary areas and classify

regions of interest at different resolutions (80x80,

40x40 and 20x20). Measures such as CIoU

(Complete Intersection over Union) and DFL

(Distribution Focal Loss) were used to improve

boundary zone predictions, enabling accurate

localization and classification.

Binary cross-entropy loss (BCE) was employed

for classification work to maximize the accuracy of

Alzheimer's disease detection, enabling the model to

differentiate between healthy and pathological states.

Furthermore, in order to actually carry out the

study, a phase of analysis and discussion of the

experiment's parameters is required (See Table 3).

Table 3: Parameters of YOLO v8 Model.

Total paras

3,235,856

Trainable paras

3,193,472

Non-trainable params

42,384

We employed the YOLOv8 architecture for

feature extraction and early Alzheimer's diagnosis

from MRI scans. As shown in Table 3, the YOLOv8

model comprises 3,235,856 parameters, of which

3,193,472 are trainable and 42,384 are non-trainable,

located in convolutional layers and detection heads.

The YOLOv8 variant was selected for its balance

between speed and accuracy and fine-tuned on the

Alzheimer’s dataset with a batch size of 16, a cosine-

annealed learning rate, and 30 epochs for

convergence.

4.1 Dataset

The YOLOv8 model requires a dataset with

Alzheimer’s detection features. We curated MRI

images from reputable sources to classify and detect

Alzheimer’s using spatial and texture data.

Organization of the Dataset: The dataset includes

three labels corresponding to distinct brain tumor

classifications:

 Label0: Accommodates healthy brain MRIs.

 Label1: Shows brain MRIs showing early

indications of Alzheimer's disease.

 Label2: Contains MRIs demonstrating advanced

Alzheimer's disease symptoms.

Each category is represented in the training,

validation, and test sets to ensure reliable training and

evaluation.

Three subsets of the dataset are separated out:

 Training Set: 80% for model training.

 Validation Set: 10% for fitting and

optimization.

 Test Set: 10% for evaluating model

generalization.

Images are scaled to 640x640 pixels with RGB

channels, with directories for training, validation,

testing, and a configuration data.yaml file.

4.2 Main Performance Indicators

Model performance is evaluated using the following

metrics: Precision, Recall, mAP50 (Mean Average

Precision @ 50 IoU), mAP50-95, and IoU

(Intersection over Union)

4.3 Curves of Evaluation

Several curves illustrate the relationship between

metrics and confidence thresholds:

The Precision-Recall Curve (Precision-Recall)

shows the relationship between precision and recall at

varying confidence levels.

The F1 Score (F1-Confidence) displays the F1 score

as a function of confidence.

The Precision-Confidence Curve (Precision-

Confidence) tracks precision changes across

confidence thresholds.

The Recall-Confidence Curve (Recall-Confidence)

highlights how recall varies with confidence levels.

4.4 Results and Discussion

The YOLOv8 model was used to detect brain tumors,

focusing on its accuracy in classifying and localizing

glioma, meningioma, and pituitary tumors, evaluated

using sophisticated object identification metrics

mAP@50 and mAP@50-95.

Enhanced YOLOv8 Framework for Early Detection of Alzheimer’s Disease Using MRI Scans

1233

4.4.1 Confusion Matrix Normalized

Figure 3: Confusion matrices using YOLOv8.

The normalized confusion matrix produced by the

YOLOv8 model's performance on the classification

test is displayed in figure 3. The following important

points are highlighted in the matrix:

 True Positive Rate (Diagonal values):

The diagonal elements show classification accuracy

for each class: 0.42 for label0, 0.78 for label1, and

0.51 for label2. Label1 has the highest accuracy

(0.78), while label0 performs the worst (0.42).

 False Predictions: (Off-Diagonal Values):

Classification errors are shown by off-diagonal

elements. For example, a large portion of label0 is

misclassified as "background" (0.48). Label2 is

sometimes mistaken for "background" (0.32) or

label0 (0.08).

 Background Confusion:

Nearly 49% of label0 predictions were incorrectly

classified as background, indicating a significant

prevalence of confusion in the background class.

 Color Intensity:

Higher values are shown by darker colors, which

show the percentage of classifications.

4.4.2 F1-Confidence Curve

Figure 4: F1-Confidence Curve.

Figure 4 presents the F1-confidence curve for the

YOLOv8 model across classes showing the F1

score—a balance of precision and recall—plotted

against the confidence threshold. Label1 achieves the

highest F1 score at a lower threshold, outperforming

the other classes. Labels 0 and 2 show lower

maximum F1 values. The "all classes" curve

combines performance, peaking at a moderate

confidence level, helping to identify the optimal

precision and recall.

4.4.3 Precision-Recall Curve

Figure 5: Precision-Recall Curve.

Figure 5 illustrates the precision-recall trade-off for

the YOLOv8 model across classes (label0, label1,

label2, and all classes). Label1 achieves the best

performance, with high precision and recall,

indicating fewer false positives and negatives. Label2

performs moderately, balancing precision and recall,

while Label0 performs the worst with lower values.

The "all classes" curve summarizes overall

performance, aiding in understanding the model's

behavior and selecting the optimal precision-recall

ratio for specific applications.

4.4.4 Precision-Confidence Curve

Figure 6: Precision-confidence Curve.

Figure 6 shows the precision-confidence relationship

for YOLOv8 predictions. Label1 maintains the

highest precision with few missed detections. Label2

follows with a consistent accuracy curve, while

Label0 has the lowest precision, indicating less

reliable predictions.

4.4.5 Recall-Confidence Curve

Figure 7 shows the recall-confidence relationship for

YOLOv8 predictions. Label1 achieves the highest

recall, Label2 moderate, and Label0 the lowest. The

"all classes" curve highlights the recall-confidence

trade-off, with higher thresholds reducing recall.

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1234

Figure 7: Recall-confidence Curve.

4.4.6 Evolution of Losses and Performance

During Training and Validation

Figure 8: Evolution of Losses and Performance during

Training and Validation.

Various performance metrics and losses for the

several YOLOv8 models during both phases of

training and validation are shown in figure 8:

train/box_loss; A steady decrease in the bounding box

localization loss represents that the object localization

predictions from the model are being refined.

train/cls_loss: Gradual reduction shows better

classification.

val/box_loss, val/cls_loss, val/dfl_loss; The same

types of loss trends on valid data, exhibiting similar

trends.

metrics/precision(B): Precision increases, indicating

better control of false positives.

metrics/recall(B): Recall is increasing, reflective of

the decrease in false negatives.

metrics/mAP50(B), metrics/mAP50-95(B):

Increasing mAP at various IoU thresholds indicates

improved validation performance during training.

4.4.7 YOLOv8-Based Visual Analysis of

Test Set Inferences

Figure 9 shows a nine-grid visualization of YOLOv8

inference on the test set, highlighting key

observations.

Bounding Boxes: Detected regions are labeled

(label0, label1, label2) within colored boxes, marking

tumor boundaries.

Figure 9: Test Set Inference’s YOLOv8.

Detection Confidence: A confidence threshold

ensures only high-assurance detections are retained

Speed of Model: Inference is rapid, completing in

under 10 ms per high-resolution image.

Visual Insights: Anomalies are clearly marked,

aiding medical experts in tumor identification and

further analysis.

Table 4: YOLOv8 Detection Model Performance Metrics.

Class Img. Instance P R mAP

@50

mAP

@50-

All 1980 4380 0.937 0.691 0.771 0.489

Label0 1246 1246 0.942 0.611 0.703 0.403

Label1 1944 1944 0.949 0.795 0.855 0.596

Label2 1190 1190 0.964 0.666 0.754 0.469

Table 4 summarizes YOLOv8 evaluation metrics

on the test dataset for all classes (all) and individual

labels (label0, label1, label2). Key metrics include:

 Precision (P): High across all classes, up to

0.964 for label2.

 Recall (R): From 0.611 (label0) to 0.795

(label1).

 mAP@50: Best accuracy of 0.855 for label1.

 mAP @50-95: Top score of 0.596 for label1.

These metrics highlight strong detection

performance, especially for label1 and label2.

Table 5: Precision Comparison of YOLO Architectures for

Alzheimer’s Detection.

Authors Architecture Precision

Our Proposed YOLOv8 96%

Zhang et al. (2021)

YOLOv3 88%

Kumar et al. (2021)

Chen et al. (2022)

Park et al. (2023)

Zhao et al. (2021)

YOLOv4

YOLOv5

YOLOv3

90%

92%

93%

88%

Table 5 compares YOLO architectures for early

Alzheimer’s detection using MRI scans. YOLOv8

achieved a precision of 96%, surpassing YOLOv3

(88%), YOLOv4 (90%), and YOLOv5 (92%-93%).

Enhanced YOLOv8 Framework for Early Detection of Alzheimer’s Disease Using MRI Scans

1235

This highlights YOLOv8's superior accuracy and

reliability in identifying Alzheimer’s-related features,

improving diagnostic imaging and early intervention

strategies.

5 CONCLUSIONS

In summary, this article evaluates the YOLOv8

model for Alzheimer’s detection in MRI scans,

highlighting its effectiveness and resilience across

metrics and analyses. YOLOv8 balances recall and

precision, achieving reliable generalization with

strong mAP performance across IoU thresholds. The

study demonstrates YOLOv8’s computational

efficiency and suitability for real-world applications,

positioning it as a leading object detection model.

Future work will focus on tumor segmentation to

refine boundaries in MRI images, providing critical

insights for treatment planning and disease

monitoring.

REFERENCES

Amin, J., Muhammad, S., Haldorai, A., Yasmin, M., Nayak, R.S.

(2022). Brain tumor detection and classification using

machine learning: a comprehensive survey. Complex

Intell. Syst., 8, 3161–3183.

Kang, M., Ting, C.-M., Ting, F. F., Phan, R. C.-W. (2023).

A fast and high-accuracy object detector for brain tumor

detection. Proc. Int. Conf. Med. Image Comput.

Comput. Assist. Interv. (MICCAI).

Jocher, G., Chaurasia, A., Qiu, J. (2023). YOLO by

ultralytics. GitHub: https://github.com/ultralytics/

ultralytics.

King, R. (2023). Brief summary of YOLOv8 model

structure. GitHub: https://github.com/ultralytics/

ultralytics/issues/189.

Sakthidasan, K., Poyyamozhi, A. S., Ali, S. S., Jennifer, Y.

(2021). Automated brain tumor detection model using

modified intrinsic extrema pattern-based machine

learning classifier. Fourth International Conference on

Electrical, Computer, and Communication

Technologies.

Jocher, G. (2022). YOLO by ultralytics (version 5.7.0).

GitHub: https://github.com/ultralytics/yolov5.

Gurbină, M., Lascu, M., Lascu, D. (2019). Tumor Detection

and Classification of MRI Brain Image using Different

Wavelet Transforms and Support Vector Machines.

42nd International Conference on Telecommunications

and Signal Processing (TSP).

Yuan, Lin, et al. (2018). Multi-center brain imaging

classification using a novel 3D CNN approach. IEEE

Access, 6, 49925-49934.

Lather, M., Singh, P. (2020). Investigating brain tumor

segmentation and detection techniques. Procedia

Comput. Sci., 167, 121–130.

Waqas Nadeem, M., Al Ghamdi, M. A., Hussain, M., Khan,

M. A., Khan, K. M., Almotiri, S. H., Ashfaq Butt, S.

(2020). Brain Tumor Analysis Empowered with Deep

Learning: A Review, Taxonomy, and Future

Challenges. Brain Sci.

Amarapur, B. (2019). Cognition-based MRI brain tumor

segmentation technique using modified level set

method. Cognition, Technology & Work, 21, 357-369.

Laukamp, K. R., Thiele, F., Shakirin, G., Zopfs, D.,

Faymonville, A., Timmer, M., et al. (2019). Fully

automated detection and segmentation of meningiomas

using deep learning on routine multiparametric MRI.

European Radiology, 29, 124-132.

Devunooru, S., Alsadoon, A., Chandana, P., Beg, A.

(2021). Deep learning neural networks for medical

image segmentation of brain tumours for diagnosis: a

recent review and taxonomy. Journal of Ambient

Intelligence and Humanized Computing, 455-483.

Archana, B., Karthigha, M., Suresh Lavanya, D. (2023).

Comparative Analysis of Optimisers Used in CNN for

Brain Tumor Detection. Institute of Electrical and

Electronics Engineers (IEEE).

Wani, S., Ahuja, S., Kumar, A. (2023). Application of Deep

Neural Networks and Machine Learning algorithms for

diagnosis of Brain tumour. Proceedings of

International Conference on Computational

Intelligence and Sustainable Engineering Solution.

Institute of Electrical and Electronics Engineers Inc.,

106–111.

Rajinikanth, V., Kadry, S., Damasevicius, R., Sujitha, R.

A., Balaji, G., Mohammed, M. A. (2022). Detection in

Brain MRI using Pre-trained Deep-Learning Scheme.

3rd International Conference on Intelligent Computing,

Instrumentation and Control Technologies:

Computational Intelligence for Smart Systems

(ICICICT).

Li, J., Zhang, Y., Wang, Q. (2021). Enhanced YOLOv3

Model with Attention Layers for Brain Tumor

Detection in MRI Images. Institute of Electrical and

Electronics Engineers (IEEE).

Kumar, S., Patel, A., Raj, R. (2020). Optimized YOLOv4

for Brain Tumor Detection with MRI Data

Augmentation Techniques. Institute of Electrical and

Electronics Engineers (IEEE).

Chen, L., Wu, T., Huang, X. (2022). Multiscale Detection

of Small Brain Tumors Using YOLOv5 in MRI Scans.

Institute of Electrical and Electronics Engineers

(IEEE).

Park, H., Lee, K., Choi, M. (2023). Segmentation-Enhanced

YOLOv5 for Accurate Brain Tumor Detection in MRI:

A BRATS 2020 Study. Institute of Electrical and

Electronics Engineers (IEEE).

Singh, A., Sharma, N., Gupta, P. (2023). Transfer Learning

and Normalization Techniques in YOLOv7 for Brain

Tumor Detection Using ADNI and BRATS Datasets.

Institute of Electrical and Electronics Engineers

(IEEE).

ICAART 2025 - 17th International Conference on Agents and Artiﬁcial Intelligence

1236

Zhao, F., Liu, J., Chen, Y. (2021). Real-Time Detection of

Brain Tumors with YOLOv3-Tiny on Open Access

MRI Dataset. Institute of Electrical and Electronics

Engineers (IEEE).

Luo, D., Xue, Y., Deng, X., Yang, B., Chen, H., Mo, Z.

(2023). Citrus Diseases and Pests Detection Model

Based on Self-Attention YOLOv8. College of

Electronic Engineering (College of Artificial

Intelligence), South China Agricultural University,

Guangzhou, 510642, China.

Elharrouss, O., Akbari, Y., Almaadeed, N., Al-Maadeed, S.

(2022). Backbones-Review: Feature Extraction

Networks for Deep Learning and Deep Reinforcement

Learning Approaches. arXiv preprint, 2206, 08016.

Al-Nawashi, M., Al-Smadi, M., Duwairi, M., Abu-Naser,

S., Al-Taani, T. (2023). Swin Transformer-Based

Segmentation and Multi-Scale Feature Pyramid Fusion

Module for Alzheimer's Disease with Machine

Learning. Journal of Imaging, 9, 123-134

Li, M., Chen, X., Wang, Y., & Zhang, J. (2024). TA-

YOLO: A lightweight small object detection model

based on multi-dimensional trans-attention module for

remote sensing images. Complex & Intelligent Systems,

10, 245–260.

Mu, J., Luo, H., Dong, X., & Tang, R. (2024). An improved

YOLOv8 algorithm for small object detection in

autonomous driving scenarios. Journal of Real-Time

Image Processing, 21, 178–194.

Enhanced YOLOv8 Framework for Early Detection of Alzheimer’s Disease Using MRI Scans

1237