INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 878

www.rsisinternational.org

An Hybrid Lightweight Model for Brain Tumor Detection

Notsa Jeff Rakotozafy*

, Andriamasinoro Rahajaniaina

, Adolphe Andriamanga Ratiarison

Department of Mathematics, Computer Science and Applications, University of Toamasina,

Toamasina, Madagascar.

Department of Physics and Applications, University of Antananarivo, Antananarivo,

Madagascar.

*Corresponding Author

DOI: https://doi.org/10.51584/IJRIAS.2025.100900087

Received: 16 September 2025 ; Accepted: 22 September 2025 ; Published: 23 October 2025

ABSTRACT

In the last decade, deep transfer learning (TL) approaches are most widely used to detect and classify

brain tumours imagines. However, current models are either complex and require significant computer

resources, or they are lightweight but use a small dataset. To overcome these problems, in this paper we

suggested a hybrid lightweight model for brain tumor detection in MRI images dataset efficiently and

accurately. Our model used MobileNetV3Small as backbone followed by a single conv layer (the neck)

to adjust the channel count and YOLO11 as detection component. So, YOLO11's inference time remains

slower than that of MobileNetV3Small. The main difficulty lies in YOLO11's feature extractor, which,

while performant, requires significant resources, limiting its use on mobile devices. To reduce

complexity and improve efficiency on mobile devices, the intermediate multi-scale head of YOLO11

(CSP/upsample fusions) is removed as it is complex. The goal is to combine the strengths of each model.

We conducted a comparative study between YOLO11 standard version and our model using the same

dataset, hyper parameters and metrics. After more experiment, the proposed model has a higher result

than YOLO11 in all metrics. It achieved 99.4% as mAP@50 and 99.8% as precision. These results have

shown that our framework is both resilient, reliable and could run on the low resource environment. For

future work, we plan to explore additional architectural optimizations and extend validation to larger,

multi-institutional datasets. Further development will focus on integrating news functionalities as

computing the dimension, area, pose of detected objects and track all mobiles objects.

Keywords: Brain tumor detection, Efficient diagnosis, Real-time object detection, lightweight model

detection

INTRODUCTION

Reference [1] described that Brain tumours stands as third most common cancer and the third leading

cause of cancer related death within adolescents and young adults. The advanced evolution of artificial

intelligence has more impacts in several domains such as medical domain. So, in their work [11]

proposed an approach to segment the images using median filter, the Otsu method is used for automated

segmentation and morphological operators for the filtering, the classification was carried out by CNN.

Reference [20] introduced ShuffleNet, a lightweight CNN to detect a brain tumor with the BraTS 2013

dataset. The accuracy of their model is reported to be 92,5%. A U-Net model using BraTS 2013 dataset

was suggested in [10], their model accomplished 93,4% as precision rate. In [7], they used SqueezeNet

with BraTS 2017 dataset for diagnosing brain tumor and achieved 94.1% of accuracy. Reference [6]

presented principal component analysis to reduce the features and used support vector machine (SVM)

for the classification of multi-sequence magnetic resonance images. In [17], they developed K-means

clustering with HSV (hue, saturation, value) colour features for the detection of the tumor and cyst using

CT image registration. A combination of three method such as fuzzy c-means, Zernike moments and

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 879

www.rsisinternational.org

region growing algorithm were developed in [13] to detect the tumor. The fuzzy c-means was used for

the MR image segmentation. Then, they applied Zernike moments to examine every tissue about the

existence or not of the tumor. At the end, they located the tumor by using region growing algorithm.

However, these previous approaches are either complex and require significant computer resources, or

they are lightweight but use a small dataset. Face to this situation, we proposed an hybrid lightweight

model has MobilNetV3Small as baseline network and YOLO11n (YOLO11 default version) as detection

component in order to overcome the previous problems. The rest of this paper is organized as follow:

brief related work on the different brain tumor detection approach is described in section II. Section III

presents materials and methods. Results and discussion about our approach are discussed in section IV

and section V conclude this study.

RELATED WORK

In this section, we discuss the existing literature about the visual method for detecting brain tumor using

images dataset. Various approaches were existed to detect the tumor. The author in [9] used YOLOv7

with transfer learning (TL) to detected early brain tumor using MRI scan. This dataset contains 3 classes

like glioma, Meningioma, and Pituitary. After experimentation, they achieved an 99,5% of accuracy.

In [14], the authors proposed a lightweight approach of diagnosing brain tumor using YOLOv5 with fine

tuning technics. They used a RSNA-MICCAI dataset partitioning to train and test. Their result gave an

accuracy 88%.

Reference [19] compared different pre-trained deep neural networks, i.e., Inceptionresnetv2,

Inceptionv3, Xception, Resnet18, Resnet50, Resnet101, Shufflenet, Densenet201 and Mobilenetv2, to

evaluate their performance in identifying and classifying different kinds of brain tumours. They added

some new layers for each pre-trained deep neural network. They employed the brain tumor classification

(MRI) dataset. Then, 80% of the data was used for training, and the remaining 20% was used for testing.

For training pre-trained DL models, they used stochastic gradient descent (SGD) through TL, 0.01

learning rate and a 10-image minibatch size. In addition, each DL model was trained for 14 epochs to

conduct the TL experiments for detecting and categorizing brain tumor types. Their experiment result

shows that the hybrid TL based on Mobilnetv2 surpassed all other ones with an accuracy rate of 82,61%.

Single Shot Detector (SSD) was developed in [3] to diagnosis a brain tumor using endoscopic images.

Their work has an accuracy of 82,7%.

The authors in [2] proposed Lightweight-CancerNet model. It is designed to detect brain tumor. Their

framework utilizes MobileNet architecture as the backbone and NanoDet as the primary detection

component. The proposed model has the ability to detect brain tumours with different image distortions.

They have used two different publicly available datasets. One is the Multimodal Brain Tumor

Segmentation Benchmark (BraTS) with 1,140 images and the other one is RSNA-MICCAI competition

data having 400 images. They used horizontal and vertical flip augmentation to expand the dataset to

1,131 images. Their result gave a mean average precision (mAP) of 93.8% and an accuracy of 98%.

A method combines Harmony Search Optimisation (HSO) and Convolution Neural Networks (CNN)

based on Deep Learning techniques proposed in [15]. They utilized CT and MRI images of brain

tumours to assess the functioning of their system. The HSO approach was utilised to extract some

information from MRI and CT images, and the CNN model was used to diagnosis brain cancers. The

result showed that their model achieved an accuracy rate of 99.13% for both detection and classification

tasks.

Reference [4] developed a method based on an improved fuzzy factor fuzzy local information C means

(IFF-FLICM) segmentation and hybrid modified harmony search and sine cosine algorithm (MHSSCA)

optimized extreme learning machine (ELM) for detecting and classifying brain tumor images. They used

image Dataset-255 for their study. The SSIM and PSNR are used for segmentation qualities measures

and sensitivity, specificity, and accuracy for detection and classification ones. The IFF-FLICM

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 880

www.rsisinternational.org

segmentation approach achieved a peak signal-to-noise ratio (PSNR) of 37.24 dB and a structural

similarity index (SSIM) of 0.9823. The MHS-SCA-based ELM model achieved a sensitivity, speci0city,

and accuracy of 98.78%, 99.23%, and 99.12%.

In [5], the authors developed a method that can automatically categorize brain tumours into various

pathological categories. The model consists of a deep neural network and an image processing

framework. It is divided into various phases, such as the mapping stage, the data augmentation stage,

and the tumor discovery stage. Their framework is based on a DCNN and a SVM algorithm to perform

generative model analysis on large datasets. The images within the dataset are equipped with a

resolution of 512 × 1024 pixels and a slice that is 6 mm thick. The data was split into a pair of datasets:

the training dataset and the test one. The first dataset was used for training the model while the testing

one performed the feature extraction. They used data augmentation technique to reduce or avoid

overfitting problem. Their model achieved a 99% accuracy rate and a sensitivity of 97.3%.

[11] presented a framework single shot multi-box detector (SSD) combined with MobileNetv2. The

proposed method used MobileNet as baseline method in order to have a lightweight model and to obtain

a best result in term of classification. The second part of their architecture is constituted by the auxiliary

network. This later was introduced for final object detection. They used the MRI image database

provided by Kaggle to train and evaluate their model. The dataset was formed by 250 MRI scans with

the infected region that is tumor. The proposed brain tumor detection method showed 98% accuracy

after 4000 epochs.

The authors in [8] introduced a method of object detection to ameliorate the accuracy of the Single Shot

Multibox Detector (SSD). They proposed the feature map. The output was obtained by transforming the

layout close to the classifier network: change the VGGNet with ResNet. However, current models are

either complex and require significant computer resources, or they are lightweight but use a small

dataset. To overcome these problems, in this paper, we presented an hybrid model for the detection of

the tumor in MRI images. The proposed model is the fusion of the MobilNetV3Small and YOLO11n.

We further incorporated the MobileNetv3Small neural network architecture within YOLO11 because it

is lightweight for the feature extraction and object classification. Our model classifies and detects the

objects faster.

MATERIALS AND METHODS

In this section, we describe all steps involved to build the proposed approach. Our method is based on

YOLO11n model combined with MobilNetV3Small as baseline network.

A- Overview and motivation

Although YOLO11n has a relatively low number of parameters (2,693,000), close to that of

MobileNetV3Small (2,542,856) in pytorch, its architecture remains complex. Furthermore, despite

improved performance compared to other YOLO versions, YOLO11's inference time remains slower

than that of MobileNetV3Small. The main difficulty lies in YOLO11's feature extractor, which, while

performant, requires significant resources, limiting its use on mobile devices.

To reduce complexity and improve efficiency on mobile devices, the goal is to combine the strengths of

each model, as MobileNetV3Small is a lightweight model capable of extracting rich features,

particularly in the C4 and C5 layers, while YOLO11 is a performant model for object detection. And

feature has the shape [batch_size, channel, height, width] where batch_size is the number of samples

processed, channel is the number of channels (or color layers), and height/width are the dimensions.

B- Design choice: replace the backbone

After conducting a comparative analysis on the features extracted from each model, we noticed that the

dimensions of the output tensors from each model are very different. MobileNetv3Small has [1, 576, 7,

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 881

www.rsisinternational.org

7] and YOLO11 has [1, 3, 640, 640]. This disparity complicates direct combination without resorting to

upsampling or downsampling operations, making the direct combination of YOLO11 and MobileNetv3

Small difficult. Therefore, we proceed by replacing YOLO11 backbone with MobileNetV3Small for

feature extraction.

To achieve this, we create within the YOLO11 network a new model YOLOmobilenetv3, which consists

of a new backbone (MobileNetV3Small), followed by a single conv layer (the neck) to adjust the

channel count, then the same Detect class used by YOLO11 for the final prediction. The intermediate

multi-scale head of YOLO11 (CSP/upsample fusions) is removed as it is complex. The structure of our

model is depicted in Fig. 1.

C- MobileNetV3Small backbone

We use depthwise separable convolutions (DWConv). Each 3×3 convolution is decomposed into a

depthwise convolution (one filter per channel, isolated) followed by a 1×1 pointwise convolution that

mixes channels; convolution is a linear operation sliding a filter over the image/features. The 3×3

convolutions (used in the backbone and head) extract local patterns. The 1×1 convolutions reduce or mix

channels without altering spatial resolution. These Conv blocks are typically followed by normalization

(BatchNorm) and activation. Batch Normalization stabilizes internal distributions and accelerates

learning. This decomposition drastically reduces parameters and computations. MobileNetV3Small also

integrates squeeze-and-excitation attention blocks and hardswish or SiLU activations for a good

accuracy/cost trade-off.

We instantiate MobileNetV3Small pretrained on ImageNet, then extract its convolutional features.

Iterating through MobileNetV3Small’s layers produce three outputs at different scales:

 p3: Output from an intermediate module (Layer 4) with ~40 channels and high resolution,

 p4: Output from layer C8 (Layer 8) with 48 channels,

 p5: Final output from layer C12 (Layer 12) with 576 channels.

Each corresponds to a different resolution (lower than the previous) and serves as a detection scale (as in

YOLO: small objects use high-resolution p3, medium objects use p4, etc.). If the highest resolution (p3)

is missing, it is recreated by upsampling p4 with bilinear interpolation. Where upsampling is a process

increasing the spatial size of a feature map, often via bilinear interpolation. In YOLO’s backbone, deep

maps are scaled for concatenation (upsampling×2). Upsampling only adds grid structure without altering

fundamental semantic information. Each output’s stride is implicit: the ratio between the input image

size and the feature map size. This operational logic is represented in Fig. 2.

D- Intermediate Neck Layer

After feature extraction, the p3, p4, p5 feature maps must be standardized for the detection head. We

therefore added a 1×1 convolution (called neck_conv) to p3, transforming its initial 24 channels to 40

channels. This pointwise convolution operation solely adjusts the channel dimension (to ensure the head

receives the expected channels) without modifying spatial size. In YOLO11's original architecture,

multiple convolutions and CSP modules were used to merge p3 with higher layers (via upsampling and

concatenation). Here, this complex fusion is simplified: we retain only a single conv1×1 layer as the

neck. This simplification avoids computational overhead while preserving sufficient spatial information

for detection. The neck acts as an intermediate network ("neck network") that prepares the backbone's

features for prediction. In many modern detectors, the neck incorporates feature pyramids (FPN/PAN,

etc.) to blend scales. Here, our neck is extremely lightweigh

(one conv layer), which is feasible because MobileNetV3Small already produces rich hierarchical

feature maps.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 882

www.rsisinternational.org

Fig. 1 architecture of the proposed model

E- Detection Head (Detect)

The one from YOLO11 remains unchanged here. It is an anchor-based implementation where each

position in a p3, p4, or p5 feature map proposes multiple candidate boxes (anchors). According to [14]

anchor-based is a strategy where box prediction relies on predefined anchors of different sizes and aspect

ratios. For each cell and each anchor, the network predicts box adjustments and class probability. This

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 883

www.rsisinternational.org

approach (popularized by Faster R-CNN and adopted in YOLOv2/v3) allows covering a range of object

dimensions. The anchors are then refined through regression.

Fig. 2 Working principle of our backbone

The Detect class uses two convolution branches for each scale: one to predict box offsets (4× reg_max

outputs per anchor) and the other for class scores (Nc outputs per anchor), then concatenates these

results. During training, we apply Distribution Focal Loss (DFL) for box regression. Specifically, instead

of directly predicting width/height, we predict a discrete distribution (with reg_max bins) for each side,

improving localization accuracy (DFL learns the exact distribution of deviations). Anchors are

dynamically generated based on each feature map's stride, then aligned with DFL predictions. Outputs

are post-processed into final boxes (center + width/height) and probability scores (sigmoid for classes).

During inference, strides (which, according to [7], are a convolution parameter defining the filter's step

size): a stride > 1 reduces resolution (downsampling). For example, a stride of 2 in 3×3 convolution

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 884

www.rsisinternational.org

doubles the step size, dividing spatial dimensions by 2. We initialized it via a dummy. Its operational

algorithm is shown in Fig. 3.

Fig. 3 Working algorithm of the detection head (Detect class)

F- Dataset Description

Glioma (a)

Meningioma (b)

pituitary tumor (c)

Fig. 4 Samples images within the dataset before pre-processing

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 885

www.rsisinternational.org

Glioma (a)

Meningioma (b)

pituitary tumor (c)

Fig. 5 Samples images within the dataset after pre-processing

Our model has used the public dataset downloaded from [20]. This brain tumor dataset containing 3064

T1-weighted contrast-inhanced images from 233 patients with three kinds of brain tumor: meningioma

(708 slices), glioma (1426 slices), and pituitary tumor (930 slices). It is a pre-labelled dataset but the

target object properties [class_id x1 y1 x2 y2] in the labels files in which class_id is a numerical class

identifier; first class has 1 as identifier. xi and yi are the planar coordinates on tumor border. These

properties not follow the YOLO format [class_id center_x center_y width height] in which class_id is a

numerical class identifier; first class has 0 as identifier. center_x and center_y are the bounding box

center coordinates (normalized 0-1). width and height are the bounding box dimensions (normalized 0-

1). To be compliance with the requirement of our model, some transformations were applied to these

ones. Thus, 90% of 3064 images were used for train set and the rest was used for validation set. Then,

some pre-processing must be applied to improve image quality. Thus, a contrast-limited adaptive

histogram equalization (CLAHE) method combined with sharpening method were used for each image.

Fig. 4 and Fig. 5 show sample images in the dataset before and after pre-processing.

Moreover, the weights of a neural network are trained using training images have a fixed size and have

the same dimensions. The images in our dataset have 640×640 as resolution. In order to conform with

the feature extractor requirement, all images must be resized into 224x224 pixels and normalized using

MobilenetV3 built-in pre-processing function before passing through our feature extractor.

RESULTS AND DISCUSSION

We conducted our experimentation on Intel(R) Core (TM) i7-1255U CPU, 10 cores, 12 threads @

2.30GHz, and 24 Gb RAM. We have been developed our system using python, torch, NumPy,

Matplotlib, torchvision, OpenCV and PIL libraries. We used precision, recall, mAP@50 and

mAP@[.50:.95] as metrics for evaluating our approach. The precision indicates the number of correct

positive predictions from the total number of actual predictions classified by the model as positive.

Recall corresponds to the score of true positive predictions to the instances that actually belong to the

positive class. mAP@50 is the means of the average precision at Intersection over Union (IoU) and

mAP@[.50:.95] indicate the mean Average Precision across all classes.

The proposed model is compared with YOLO11n using the same dataset. A lot of experiments were

conducted during the training phase but the batch size of 8 and an epoch of 700 performed the best

result. The Fig. 6 shows the curves of results.

From plot is seen that our model achieved the best result than YOLO11 model.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 886

www.rsisinternational.org

Fig. 6.a F1-Confidence Curve of YOLO11

Fig. 6.b Precision-Recall Curve of YOLO11

Fig. 6.c F1-Confidence Curve of our Model

Fig. 6.d Precision-Recall Curve of our Model

To evaluate the performance of our YOLOv11-MobileNetV3-small, we first tested it on PyTorch using an

inference script, meaning we performed detection on multiple images in a folder. The results are shown in

Figure 7 below.

Fig. 7.a Detection of glioma

Fig. 7.b Detection of meningioma

Fig. 7.c Detection of pituitary tumor

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 887

www.rsisinternational.org

Then, we converted the model to TensorFlow Lite for use on a mobile platform. We imported the YOLOv11-

MobileNetV3-small TFLite model into a mobile application we designed and performed detection on the

mobile device, where the detection results along with the inference time are presented in Figure 8.

Fig. 8.a Detection of glioma on android device

Fig. 8.b Detection of meningioma on android device

Fig. 8.c Detection of pituitary on android device

Table I below shows that the proposed model demonstrates higher performance than YOLO11 using the same

dataset, obtaining a precision of 99.8%, a recall of 98.8%, a mAP@50 of 99.4% and mAP@ [.50:.95] of 98.4

%. Significantly, this surpasses all results attained by YOLO11n.

Table I: comparative results

Model

Precision

Recall

mAP@50

mAP@[.50:.95

YOLO11n (default)

83%

77.4%

84.1%

57.4%

Proposed Model

99.8%

98.8%

99.4%

98.4%

Figure 7 shows sample detections produced by our model.

CONCLUSION

In this work, we proposed a hybrid lightweight model for brain tumor detection by integrating

MobileNetV3Small as a feature extraction with the YOLO11 detection framework. The model was rigorously

evaluated and compared against the standard YOLO11 implementation across 700 training epochs. Our

proposed architecture demonstrated superior performance across all key metrics, achieving a precision of

99.8% (compared to 83% for YOLO11), recall of 98.8% (versus 77.4%), mAP@50 of 99.4% (versus 84.1%),

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 888

www.rsisinternational.org

and mAP@50-95 of 98.4% (versus 57.4%). These results represent a significant improvement in detection

accuracy while maintaining computational efficiency suitable for low-resource environments.

The substantial performance gains confirm the effectiveness of our architectural modifications, particularly the

replacement of the backbone and simplification of the neck network. This approach successfully reduces

computational complexity while enhancing feature extraction capabilities for medical imaging applications.

For future work, we plan to explore additional architectural optimizations and extend validation to larger,

multi-institutional datasets. Further development will focus on using this model in new datasets.

REFERENCES

1. Aaron Cohen-Gadol, (2024). Must-known brain tumor statistics. www.aaroncohen-gordol.com.

2. Asif Raza and Muhammad Javed Iqbal, (2025). Lightweight-CancerNet: a deep learning approach for

brain tumor detection. In PeerJ Computer Science 11:e2670 DOI 10.7717/peerj-cs.2670.

3. Cen Q, Pan Z, Li Y, Ding H., (2019). Laryngeal tumor detection in endoscopic images based on

convolutional neural network. In: IEEE 2nd International Conference on Electronic Information and

Communication Technology. Piscataway: IEEE.

4. Dash et al, (2024). Brain Tumor Detection and Classification Using IFF-FLICM Segmentation and

Optimized ELM Model, Journal of Engineering Volume 2024, Article ID 8419540, 24 pages

https://doi.org/10.1155/2024/8419540.

5. Gunasundari and Selva Bhuvaneswari, (2025). A novel approach for the detection of brain tumor and

its classification via independent component analysis. Scientific reports. 14 pages.

https://doi.org/10.1038/s41598-025-87934-4.

6. Gupta et al., (2005). Support vector machine for optical diagnosis of cancer. Journal of Biomedical

Optics 10(2), 024034.

7. Goodfellow, I., et al., (2016). Deep Learning. MIT Press.

8. Iandola FN, et al., (2016). SqueezeNet: a simple and efficient CNN for image classification. In: IEEE

Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 3511–3519.

9. J. Jeong, H. et al., (2017). Enhancement of SSD by concatenating feature maps for object detection.

arXiv preprint arXiv:1705.09587.

10. Kumar A., (2023). Study and analysis of different segmentation methods for brain tumor MRI

application. Multimedia Tools and Applications 82(5):7117–7139. DOI 10.1007/s11042-022-13636-.

11. Menze BH, et al., (2015). The multimodal brain tumor image segmentation benchmark (BRATS).

IEEE Transactions on Medical Imaging 34(10):1993–2024 DOI 10.1109/TMI.2014.2377694.

12. Naseer-u-Din et al. (2022). Brain tumor detection in MRI scans using single shot multibox detector. in

Journal of Intelligent & Fuzzy Systems · March 2022 DOI: 10.3233/JIFS-219298.

13. P. K. Barik and R. C., (2019). Barik. Feed forwarded ct image registration for tumour and cyst

detection using rigid transformation with hsv colour segmentation. International Journal of

Computational Systems Engineering, 5(5-6):277–286.

14. P. Maji and S. Roy. SoBT-RFW, (2015). rough-fuzzy computing and wavelet analysis based automatic

brain tumor detection method from MR images. Fundamenta Informaticae, 142(1-4):237–267.

15. Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-CNN: Towards real-time object detection

with region proposal networks. Advances in Neural Information Processing Systems, 28, 91–99

16. Shelatkar T, et al., (2022). Diagnosis of brain tumor using light weight deep learning model with fine-

tuning approach. Computational and Mathematical Methods in Medicine 2022:1–9 DOI

10.1155/2022/2858845.

17. Shaik et al., (2024). Detection and Classification of Brain Tumor from MRI And CT Images using

Harmony Search Optimization and Deep Learning. Journal of Artificial Intelligence Research &

Advances, ISSN: 2395-6720 Volume 11, Issue 3, 2024 September–December DOI Journal:

10.37591/JoAIRA.

18. S. Naz and N. Kumar, (2019). An Efficient Brain Tumor Detection system using Automatic

segmentation with Convolution Neural Network.

INTERNATIONAL JOURNAL OF RESEARCH AND INNOVATION IN APPLIED SCIENCE (IJRIAS)

ISSN No. 2454-6194 | DOI: 10.51584/IJRIAS |Volume X Issue X October 2025

Page 889

www.rsisinternational.org

19. Shahariar et al., (2019). Automatic Human Brain Tumor Detection in MRI Image Using Template-

Based K Means and Improved Fuzzy C Means Clustering Algorithm. In Big Data Cogn. Comput.

20. Sorour, A. (2024). MRI for Brain Tumor with Bounding Boxes. Kaggle.

https://www.kaggle.com/datasets/ahmedsorour1/mri-for-brain-tumor-with-bounding-boxes

21. T. Gupta, et al., (2017). Multi-sequential mr brain image classification for tumor detection. Journal of

Intelligent & Fuzzy Systems, 32(5):3575–3583.

22. Ullah N, et al., (2022). An effective approach to detect and identify brain tumours using transfer

learning. Applied Sciences 12(11):5645 DOI 10.3390/app12115645.

23. Zhang et al. (2017). ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile

Devices. arXiv:1707.01083. https://doi.org/10.48550/arXiv.1707.01083