2023 4th International Conference on Intelligent Technologies (CONIT)

Karnataka, India. Jun 21-23, 2024

Hybrid Deep Transfer Learning Framework for

Humerus Fracture Detection and Classification from
X-ray Images
Puja Dey Tanjim Mahmud Khan Md. Foysol
Department of CSE Department of CSE Dept. of EEE
University of Chittagong Rangamati Science and Textile Engineering College, Noakhali
Chittagong -4331,Bangladesh Technology University Chowmuhani 3821, Bangladesh
[email protected] Rangamati-4500, Bangladesh [email protected]
tanjim [email protected]

Nahed Sharmen Mohammad Shahadat Hossain Karl Andersson

Dept. of Applied Microbiology Department of CSE Cybersecurity Laboratory
Kitami Institute of Technology University of Chittagong Luleå University of Technology
Hokkaido 090-0015, Japan Chittagong -4331,Bangladesh 97187 Luleå, Sweden
[email protected] hossain [email protected] [email protected]

Abstract—The detection and classification of humerus frac- treatment. Traditional methods of fracture detection through
tures from X-ray images are crucial for effective medical diagno- manual examination of X-ray images are laborious, time-
sis and treatment planning. Manual assessment of such fractures consuming, and prone to error [3]. In response, there’s a grow-
is time-consuming and prone to errors, emphasizing the need
for automated systems. In this study, we propose a Hybrid Deep ing imperative for automated systems capable of swiftly and
Transfer Learning Framework for Humerus Fracture Detection accurately identifying fractures, thereby expediting diagnosis
and Classification from X-ray Images. Leveraging deep learning and treatment planning.
techniques, we amassed a dataset of 1266 radiographic images In this research, we present an approach leveraging deep
from the publicly available MURA dataset, encompassing both learning techniques for the automated detection and classifica-
negative (non-fractured) and positive (fractured) cases. Prepro-
cessing techniques were employed to enhance image quality, tion of humerus fractures from X-ray images. We assembled a
followed by data augmentation to mitigate overfitting and bol- dataset comprising 1266 radiographic images of the humerus
ster system accuracy. Subsequently, a hybrid model comprising bone, encompassing both negative (non-fractured) and pos-
ResNet50 and DenseNet121 architectures was utilized for fea- itive (fractured) instances, sourced from the publicly avail-
ture extraction and classification. Through experimentation with able ”MURA” dataset [4]. Preprocessing methodologies were
various optimizers, we achieved the highest accuracy of 93.41%
using the Adam optimizer. Additionally, precision, recall, and applied to enhance image quality and facilitate subsequent
F1-score metrics were computed to evaluate model performance analysis, complemented by data augmentation techniques to
comprehensively. Comparative analyses were conducted with mitigate overfitting and bolster system accuracy [5]. Central to
other pre-trained models, showcasing the effectiveness of our our methodology is a hybrid deep transfer learning framework,
proposed framework. Our results highlight the deep transfer amalgamating the strengths of two state-of-the-art convolu-
learning’s effectiveness in humerus fracture detection, providing
a promising path forward for the development of medical imaging tional neural networks (CNNs) [5]–[7], namely ResNet50 and
technologies. DenseNet121 [8]. These networks are employed for feature
Index Terms—Deep Learning, CLAHE, Hybrid Model, Trans- extraction and classification, discerning between non-fractured
fer Learning, Augmentation and fractured humerus instances. Various optimizers [6], [9]–
[12] were explored, with the Adam optimizer yielding the
I. I NTRODUCTION highest accuracy of 93.41%.
The skeletal system is essential to human physiology be- In addition, we carried out a thorough analysis, calculating
cause it consists of bones that support and shield important metrics such as precision, recall, and F1-score [13] to gauge
organs structurally [1]. The humerus is one of these bones; the model’s effectiveness. We compared our hybrid model
it is located in the upper arm, between the shoulder and the to other pre-trained models and previous research to put our
elbow, and it is essential for supporting upper body functions findings in context and to show the effectiveness and future
and enabling arm movement [2]. The integrity of the humerus potential improvements provided by our suggested framework.
is damaged by fractures, which greatly interferes with day- By providing a reliable and accurate method for humerus
to-day activities and calls for prompt, precise diagnosis and fracture detection through this work, we hope to support the

979-8-3503-4990-0/24/$31.00 ©2024 IEEE 1

ongoing efforts in medical image analysis and potentially accuracy of 88.5% and 90.94% respectively from SENet154,
improve patient care and clinical decision-making. and DenseNet201 model. For classification and abnormity
The contributions of the paper are below: detection of bone radiographic images El-Saadawy [22] intro-
1) Utilizing deep Convolutional Neural Network for detect- duced a hybrid GNG modified VGG approach which consisted
ing humerus fracture. of two stages. In the second stage they obtained sensitivity
2) Applying CLAHE (contrast-limited adaptive histogram (Recall) of 92.50% for detecting bone deformation. Sezer
equalization) for enhancing image quality. et al. [23] trained segmented shoulder images for extracting
3) Implementing several pre-trained models (ResNet50, features and classifying the humerus head into three types:
VGG19, VGG16, DenseNet121) with transfer learning. normal, edematous, and Hill-Sachs lesion utilizing CNN. They
4) Introducing hybrid model with best testing accuracy of obtained 98.43% accuracy.
93.41%
III. M ETHODOLOGICAL F RAMEWORK
II. P RIOR S TUDIES E XAMINATION In this section we illustrate the whole process of the system
In this section, there is an overview of existing research which has been implemented in our research of humerus
work related to bone fracture detection. Chawla et al. [2] in fracture detection. As mentioned in figure 1, we started our
2020 developed a ensemble model combining DenseNet201 study by collecting dataset. Then some pre-processing and data
and Inception V3 to classify humerus bone fracture using augmentation techniques [24]–[26] were applied for enhanc-
’MURA’ dataset and obtained accuracy of 88.54% with F1- ing performance. After that these preprocessed images were
score of 0.892. In 2023, Sumi et al. [14] performed the passed through an hybrid model for feature extraction. And
implementation using a CNN model as well as some pre- finally, the classification was done to categorize the data as
trained model. Their best performing model is CNN with ”Negative”(non-fractured) or ”Positive” (fractured).
testing accuracy of 78%. Sashidhar et al. [15], in their re-
search utilised three pre-trained model (VGG16, DenseNet121, A. Dataset
DenseNet169) and achieved promising result with 80% ac- For both the training and testing phases, we used the
curacy using DenseNet121. Mondol et al. [16] introduced MURA dataset [4] of musculoskeletal abnormalities. It is one
a Computer Aided Diagnosis (CADx) model using Deep of the biggest radiograph datasets that is freely accessible.
convolutional neural network to detect bone abnormalities. The Stanford University Machine Learning Group published
They applied ResNet and VGG19 models and combined these it after certified radiologists from Stanford Hospital annotated
two models to develop an ensemble model named CADx it. The dataset includes 40,895 radiographic pictures of 12,251
model. This CADx model obtained comparatively better result patients’ musculoskeletal conditions. The seven different bone
(87.15% for humerus classification) than both ResNet and types found in this dataset are the elbow, finger, forearm,
VGG19. A new model named capsule network was developed hand, humerus, shoulder, and wrist. Every category features
by Saif et al. [3]. Cheng et al. [17] employed adversarial policy a binary indicator that indicates whether the bone is fractured
gradient augmentation (APGA) for carring out classification (positive) or not (negative). However, the primary focus of our
utilizing the dataset consisting of hip bone images and the investigation was humerus fracture detection. There are 1,266
whole MURA dataset. They obtained best accuracy of 86.53% humeral radiographs from 727 patients total; 669 of these are
for humerus images detection. Galal et al. [18] utilised a negative cases and 597 are positive cases.
small part of elbow radiographic images (train images: 56, Figure 2 shows some images from ’MURA’ dataset.
test images: 24) from MURA datset for implementation. They
achieved 97% accuracy using SVM and 91.6% accuracy using B. Pre-processing
both random forest and naive bayes. Rajpurkar et al. [4] Image pre-processing techniques help in noise reduction,
implemented a DenseNet169 model to classify and detect bone enhanced feature extraction, and intensity normalization of
deformation using MURA dataset. Utilising this model they images [27]. We resized all the radiographic images to a
got sensitivity of 0.815 and specificity of 0.887. Thian et dimension of 96*96 so that the system can assure coherent
al. [19] adopted a model for determining the feasibility and feature extraction, reduced processing time, and finally get
performance of Convolutional Neural Network when detecting an improved performance. CLAHE (contrast-limited adaptive
and localizing fracture from x-ray images. Their final model histogram equalization) method has been employed over the
was developed utilising Inception-ResNet and a Faster R-CNN resized images. For normalization the equalized images have
for detecting fracture of wrist images. Pelka et al. introduced been converted into array of pixel value. Then each pixel value
a LSTM based RNN model. Using whole MURA dataset, has been divided by 255 for rescaling the pixel values to range
they achieved 95.93% and 81.5% accuracy respectively for between 0 to 1. Figure 3 shows the pre-processing steps of our
anatomic regions and abnormality prediction. Pradhan et al. system.
[20] trained a deep CNN to classify the whole MURA dataset
and obtained an accuracy of 91.37%. Shao and Wang [21] C. Data Augmentation
proposed a two stage approach to train the whole MURA Data augmentation comprises methods to enlarge the ex-
dataset for abnormity detection. For humerus images they got isting dataset, which helps improving classification accuracy

2
 Fig. 1: Workflow diagram of our proposed work

E. Transfer Learning Framework

To detect and classify fracture of humerus, we applied trans-
fer learning with four remarkable network [30] - ResNet50,
VGG19, VGG16, and DenseNet121. Dataset fine tuning neces-
sitated adjusting fully connected layers while preserving pre-
trained convolution layers. The final prediction layer has been
modified for binary classification. This method quickened the
Fig. 2: Sample dataset training process and improved the performance of our system
by adjusting models to the features of dataset. Utilization of
transfer learning in pre-trained models affirmed efficiency in
enhancing accuracy and credibility for detecting the fractures
in radiographic images.
1) ResNet-50: This Residual Convolutional Neural Net-
work contains 50 layers, of which the first layer is a convolu-
tion layer of 64 filters with 7*7 filter size, and stride value of
Fig. 3: Pre-processing steps 2. The next three stages comprise of convolution layers with
1*1, 3*3, 1*1 kernel sizes, each having 64, 64, and 256 filters,
respectively. These stages are repeated three times. Similarly,
and alleviating overfitting [28]. In our study there is a small the subsequent convolution layers are repeated four times, and
dataset of 1266 x-ray images of humerus bone fracture so we the following convolution blocks are repeated six times.
applied data augmentation for bolstering the stability of neural 2) VGG16: The input image dimensions in this network
network. For this purpose, the ImageDataGenerator has been are 224*224*3. The first two convolution layers, which have
employed. Following table I shows the augmentation settings. 64 filters each and a kernel size of 3*3, are followed by a
max pooling layer that has a pool size of 2*2 and a stride
TABLE I: Data Augmentation Setting value of 2. The next two convolution layers have a kernel size
Augment techniques Range of 3 × 3 and contain 128 filters, after which there is another
Rotation 40 max pooling layer. Three convolution layers with 256 filters
Shear Range 0.2 and a 3*3 kernel size follow this pooling layer. The final max-
Zoom Range 0.15
Horizontal flip True pooling layer and two sets of three convolution layers with 512
filters and a 3*3 kernel size follow these convolution layers.
In the end, two dense layers are added for classification along
with one flat layer.
D. Feature Extraction
3) VGG19: The input images in this network have the same
For feature extraction, we applied a pre-trained version dimensions, 224*224*3. It also has 64 features in the first
of RestNet50 model. It has been employed because of it’s two convolution layers, each with a 3*3 kernel size. The max
structure and depth. Before passing the radiographic images pooling layer comes next, with a 2*2 pool size and a stride
through the ResNet50 model some preprocessing techniques value of 2. The following two convolution layers, each with a
have been utilized to assure that these input images are in 3 × 3 kernel, carry 128 features, and are followed by a second
proper format. When X-ray images are forwarded through max pooling layer. Four convolution layers with 256 features
network, it extracts features at various levels of abstraction. and a 3*3 kernel size follow this pooling layer. The max-
This process of feature extraction begins with extracting basic pooling layer comes after two sets of four convolution layers
characteristics such as textures and edges in initial layer and with 512 features and a 3*3 kernel size. Additionally, two
advancing to more discriminatory and intricate characteristics dense layers and one flatten layer are present for categorization
in deeper layer [29]. purposes.

3
 4) DenseNet121: Four dense blocks are carried by TABLE III: Comparison of results among different optimizers
DenseNet121. Using kernel sizes of 1*1 and 3*3, convolution based on accuracy
is achieved in the first dense block. This step is repeated six Models Train Acc Test Acc Train Loss Test Loss
times. Additionally, convolution is carried out with 3*3 and Adam 99.23% 93.41% .1231 .2543
Adamax 96.54% 91.10% .1539 .2831
1*1 kernel sizes in the second dense block, and it is done 12 Adagrad 94.02% 89.96% .1805 .2984
times. Convolution operations are repeated 24 and 16 times, SGD 90.49% 83.68% .2191 .3204
respectively, with the same kernel size in the third and fourth Adadelta 87.79% 81.86% .2435 .3386
dense blocks. Between dense blocks are transition blocks made
up of pooling and convolution layers.
compared to the true result. On the other hand, the losses are
F. Hybrid Model determined by the summation of the mistakes created from
Hybrid structure has the ability of combining various base each sample of training or testing data.
estimators or models to increase the classification accuracy.
Therefore, in our research, We implemented a hybrid model
to enhance the efficiency and reliability of our approach.
After the implementation of all transfer learning based pre-
trained models, we achieved a highest accuracy of 85.79%.
Further, for improving accuracy, we introduced a hybrid model
combining ResNet50, and DenseNet121 along with some
additional layers. Table II illustrates the description of our
proposed model.

TABLE II: Hyperparameters Setting

Content Details
Input image size 224*224*3
ResNet50(Sequential 1) a) Accuracy
DenseNet121(Sequential
2)
1st Conv. Layer filters=32,kernel=3*3,ReLu,padding=same
1st Maxpool Layer pool size = 2*2
2nd Conv. Layer filters=64,kernel=3*3,ReLu,padding=same
2nd Maxpool Layer pool size = 2*2
3rd Conv. Layer filters=128,kernel=3*3,ReLu,padding=same
3rd Maxpool Layer pool size = 2*2
Dropout Layer 50% neurons dropped
1st Dense layer 256 nodes, ReLu
2nd Dense Layer 128 nodes, ReLu
Output Layer 2 nodes, Sigmoid activation
Optimizer Adam
Learning rate .001
Loss Function Binary Crossentropy

b) Loss
IV. F INDINGS AND I NTERPRETATION
Fig. 4: Accuracy and Loss Curves of proposed model
A. Performance Metrics
To evaluate the efficiency and calculate the overall perfor- The X axle in Figure 4 above represents the number of
mance of our suggested hybrid model, we have determined training and testing samples for each epoch, while the Y
several matrices based on performance [1], [31], [32] i.e. F1- axle represents the accuracy and loss, respectively. Figure
score, Precision, and Recall. 4(a) illustrates how, after 20 epochs, the testing and training
accuracies increase from 19.03% to 93.41% and 16.67%
B. Classification results of the proposed architectures to 99.23%, respectively. Furthermore, after 20 epochs, the
In this subsection we illustrates the classification results of training and testing losses in figure 4(b) decrease from 2.52 to
humerus fracture detection. Following table III demonstrates 0.12 and 2.25 to 0.25, respectively. The precision, recall, and
the training and testing accuracies and losses of our dataset F1-score score of the suggested model are displayed in Table
using different optimizers. We have utilised Adam, Adamax, IV.
Adagrad, SGD and Adadelta optimizers in our system and
achieved best accuracy using Adam optimizer. C. Comparison of proposed model
Figure 4 shows the accuracy and loss curves of training Four pre-trained models based on transfer learning were
and testing data using our proposed hybrid model. Where used in our study to train the dataset so that we could
accuracies identify how perfectly a model can classify data compare our recommended method with other pre-trained

4
TABLE IV: Precision, Recall and F1-score Score of Proposed TABLE VI: Comparison with State-of-art Method
Model Reference Model Results
Performance Negative Positive Macro [1] Ensemble Accuracy: 85% ,F1-score: 0.79
Measures Average [14] CNN Accuracy: 78%
Precision 0.92 0.95 0.93 [15] DenseNet121 Accuracy: 80%
Recall 0.95 0.90 0.92 [2] Ensemble Accuracy: 88.54% , F1-score:
F1-score 0.93 0.92 0.92 0.892
Our Hybrid Accuracy: 93.41%, F1-score: 0.93
Study
models. We have employed the ResNet50, VGG19, VGG16,
and DenseNet121 models in relation to this. We have discov-
ered that our recommended hybrid model is outperforming V. C ONCLUSION AND F UTURE WORK
other models after putting these algorithms into practice. In this paper, we introduced a hybrid deep transfer learning
The confusion matrix for our suggested model and other framework for humerus fracture detection and classification
pre-trained models are displayed in Figure 5. Additionally, from X-ray Images. By utilizing machine learning and com-
Table V showed how various parameters from each model puter vision techniques, we successfully tackled the crucial
were compared. problem of automating the identification and categorization of
humerus fractures. This represents a noteworthy advancement
in improving accuracy in diagnosis and patient care within
the medical field. To enhance our model’s performance, we
used a variety of preprocessing methods in our study, such as
Contrast Limited Adaptive Histogram Equalization (CLAHE).
In addition, we put forth a hybrid model that combined
a) Resnet50 b) DenseNet121 ResNet50, DenseNet121, and additional layers, attaining re-
markable outcomes with 93.41% accuracy, 0.9355 precision,
0.9324 recall, and 0.9339 F1-score. The comparative analysis
demonstrated our proposed model’s superiority over a number
of pretrained models based on transfer learning, including
ResNet50, VGG19, VGG16, and DenseNet121.
Our model demonstrated superior performance metrics, val-
c) VGG19 d) VGG16 idating its efficacy in humerus fracture detection and classifi-
cation tasks. Additionally, comparing our work with previous
studies highlighted the advancements achieved in our system,
further emphasizing the significance of our contributions.
The dataset will be diversified in the future, along with other
optimization techniques [33]–[41] and model architectures,
and the framework will be validated in clinical settings.
e) Hybrid For practical application, special attention will be paid to
improving interpretability and incorporating the model into
Fig. 5: Confusion matrices (0-negative, 1-positive) diagnostic workflows.

TABLE V: Comparison of results among different models R EFERENCES

based on several performance parameters [1] K. Barua, T. Mahmud, A. Barua, N. Sharmen, N. Basnin, D. Islam, M. S.
Model Accuracy Precision Recall F1-score Hossain, K. Andersson, and S. Hossain, “Explainable ai-based humerus
fracture detection and classification from x-ray images,” in 2023 26th
ResNet50 85.79% 0.8572 0.8555 0.8563
International Conference on Computer and Information Technology
DenseNet121 85.09% 0.8507 0.8492 0.8499
(ICCIT). IEEE, 2023, pp. 1–6.
VGG19 84.32% 0.8419 0.8406 0.8412
VGG16 83.78% 0.8357 0.8342 0.8349 [2] N. Chawla and N. Kapoor, “Musculoskeletal abnormality detection in
Hybrid Model 93.41% 0.9355 0.9324 0.9339 humerus radiographs using deep learning.” Rev. d’Intelligence Artif.,
vol. 34, no. 2, pp. 209–214, 2020.
[3] A. Saif, C. Shahnaz, W.-P. Zhu, and M. O. Ahmad, “Abnormality
In table VI, we have illustrated a comparison of our pro- detection in musculoskeletal radiographs using capsule network,” IEEE
posed hybrid approach with some existing research which Access, vol. 7, pp. 81 494–81 503, 2019.
have done on this topic. Here, we mentioned the methods that [4] P. Rajpurkar, J. Irvin, A. Bagul, D. Ding, T. Duan, H. Mehta, B. Yang,
K. Zhu, D. Laird, R. L. Ball et al., “Mura: Large dataset for ab-
have been used in those papers. We also mentioned about the normality detection in musculoskeletal radiographs,” arXiv preprint
accuracies that they obtained using their model. Furthermore, arXiv:1712.06957, 2017.
we observe that our suggested approach is superior to other [5] T. Mahmud, K. Barua, A. Barua, N. Basnin, S. Das, M. S. Hossain, and
K. Andersson, “Explainable ai for tomato leaf disease detection: Insights
approaches with a 93.41% accuracy rate in detecting humerus into model interpretability,” in 2023 26th International Conference on
fractures. Computer and Information Technology (ICCIT). IEEE, 2023, pp. 1–6.

5
 [6] M. H. Imam, N. Nahar, R. Bhowmik, S. B. S. Omit, T. Mahmud, [24] D. Islam, T. Mahmud, and T. Chowdhury, “An efficient automated vehi-
M. S. Hossain, and K. Andersson, “A transfer learning-based framework: cle license plate recognition system under image processing,” Indonesian
Mobilenet-svm for efficient tomato leaf disease classification,” in 2024 Journal of Electrical Engineering and Computer Science, vol. 29, no. 2,
6th International Conference on Electrical Engineering and Information pp. 1055–1062, 2023.
& Communication Technology (ICEEICT), 2024, pp. 693–698. [25] T. Mahmud, A. Barua, M. Begum, E. Chakma, S. Das, and N. Sharmen,
[7] T. Mahmud, M. Ptaszynski, and F. Masui, “Deep learning hybrid “An improved framework for reliable cardiovascular disease prediction
models for multilingual cyberbullying detection: Insights from bangla using hybrid ensemble learning,” in 2023 International Conference on
and chittagonian languages,” in 2023 26th International Conference on Electrical, Computer and Communication Engineering (ECCE). IEEE,
Computer and Information Technology (ICCIT). IEEE, 2023, pp. 1–6. 2023, pp. 1–6.
[8] T. Mahmud, A. Barua, K. Barua, N. Basnin, M. S. Hossain, K. An- [26] N. S. Nuha, T. Mahmud, N. Rezaoana, M. S. Hossain, and K. Andersson,
dersson, M. S. Kaiser, M. S. Hossain, M. Monju, and N. Sharmen, “An approach of analyzing classroom student engagement in multimodal
“Enhancing diagnosis: An ensemble deep learning model for brain tumor environment by using deep learning,” in 2023 IEEE 9th International
detection and classification,” in International Conference on Big Data, Women in Engineering (WIE) Conference on Electrical and Computer
IoT and Machine Learning. Springer, 2023, pp. 409–424. Engineering (WIECON-ECE). IEEE, 2023, pp. 286–291.
[9] N. A. Chowdhury, T. Mahmud, A. Barua, N. Basnin, K. Barua, A. Iqbal, [27] T. Mahmud, S. Das, M. Ptaszynski, M. S. Hossain, K. Andersson,
M. S. Hossain, K. Andersson, M. S. Kaiser, M. S. Hossain et al., “A and K. Barua, “Reason based machine learning approach to detect
novel approach to detect stroke from 2d images using deep learning,” bangla abusive social media comments,” in International Conference on
in International Conference on Big Data, IoT and Machine Learning. Intelligent Computing & Optimization. Springer, 2022, pp. 489–498.
Springer, 2023, pp. 239–253. [28] S. Das, T. Mahmud, D. Islam, M. Begum, A. Barua, M. Tarek Aziz,
[10] T. Mahmud, I. Hasan, M. T. Aziz, T. Rahman, M. S. Hossain, and E. Nur Showan, L. Dey, E. Chakma et al., “Deep transfer learning-based
K. Andersson, “Enhanced fake news detection through the fusion of foot no-ball detection in live cricket match,” Computational Intelligence
deep learning and repeat vector representations,” in 2024 2nd Interna- and Neuroscience, vol. 2023, 2023.
tional Conference on Intelligent Data Communication Technologies and [29] T. Mahmud, K. Barua, A. Barua, S. Das, N. Basnin, M. S. Hos-
Internet of Things (IDCIoT). IEEE, 2024, pp. 654–660. sain, K. Andersson, M. S. Kaiser, and N. Sharmen, “Exploring deep
[11] T. Mahmud, M. Ptaszynski, and F. Masui, “Automatic vulgar word transfer learning ensemble for improved diagnosis and classification of
extraction method with application to vulgar remark detection in chit- alzheimer’s disease,” in International Conference on Brain Informatics.
tagonian dialect of bangla,” Applied Sciences, vol. 13, no. 21, p. 11875, Springer, 2023, pp. 109–120.
2023. [30] T. Mahmud, K. Barua, S. U. Habiba, N. Sharmen, M. S. Hossain, and
K. Andersson, “An explainable ai paradigm for alzheimer’s diagnosis
[12] F. Khan, R. Mustafa, F. Tasnim, T. Mahmud, M. S. Hossain, and
using deep transfer learning,” Diagnostics, vol. 14, no. 3, p. 345, 2024.
K. Andersson, “Exploring bert and elmo for bangla spam sms dataset
[31] P. Dey, T. Mahmud, S. R. Nahar, M. S. Hossain, and K. Andersson,
creation and detection,” in 2023 26th International Conference on
“Plant disease detection in precision agriculture: Deep learning ap-
Computer and Information Technology (ICCIT). IEEE, 2023, pp. 1–6.
proaches,” in 2024 2nd International Conference on Intelligent Data
[13] T. Mahmud, M. Ptaszynski, J. Eronen, and F. Masui, “Cyberbullying
Communication Technologies and Internet of Things (IDCIoT). IEEE,
detection for low-resource languages and dialects: Review of the state
2024, pp. 661–667.
of the art,” Information Processing & Management, vol. 60, no. 5, p.
[32] T. Mahmud, M. Ptaszynski, and F. Masui, “Exhaustive study into ma-
103454, 2023.
chine learning and deep learning methods for multilingual cyberbullying
[14] T. A. Sumi, N. Basnin, M. S. Hossain, K. Andersson, and M. S. detection in bangla and chittagonian texts,” Electronics, vol. 13, no. 9,
Hoassain, “Classifying humerus fracture using x-ray images,” in The p. 1677, 2024.
Fourth Industrial Revolution and Beyond: Select Proceedings of IC4IR+. [33] M. M. Islam, T. Mahmud, and M. S. Hossain, “Belief-rule-based intel-
Springer, 2023, pp. 527–538. ligent decision system to select hospital location,” Indonesian Journal
[15] A. Sasidhar, M. Thanabal, and P. Ramya, “Efficient transfer learning of Electrical Engineering and Computer Science, vol. 1, no. 3, pp. 607–
model for humerus bone fracture detection,” Annals of the Romanian 618, 2016.
Society for Cell Biology, pp. 3932–3942, 2021. [34] T. Mahmud and M. S. Hossain, “An evidential reasoning-based decision
[16] T. C. Mondol, H. Iqbal, and M. Hashem, “Deep cnn-based ensemble support system to support house hunting,” International Journal of
cadx model for musculoskeletal abnormality detection from radio- Computer Applications, vol. 57, no. 21, pp. 51–58, 2012.
graphs,” in 2019 5th International Conference on Advances in Electrical [35] R. Karim, M. Khaliluzzaman, T. Mahmud et al., “An expert system for
Engineering (ICAEE). IEEE, 2019, pp. 392–397. clinical risk assessment of polycystic ovary syndrome under uncertainty,”
[17] K. Cheng, C. Iriondo, F. Calivá, J. Krogue, S. Majumdar, and V. Pedoia, 2023.
“Adversarial policy gradient for deep learning image augmentation,” in [36] T. Mahmud, K. N. Rahman, and M. S. Hossain, “Evaluation of job offers
Medical Image Computing and Computer Assisted Intervention–MICCAI using the evidential reasoning approach,” Global Journal of Computer
2019: 22nd International Conference, Shenzhen, China, October 13–17, Science and Technology, vol. 13, no. 2, pp. 34–44, 2013.
2019, Proceedings, Part VI 22. Springer, 2019, pp. 450–458. [37] J. Sikder, T. Mahmud, B. Banik, and S. Gupta, “Linear programming to
[18] A. Galal, F. Hisham, M. Mohamed, S. Hassan, T. Ghanim, and A. Nabil, find the critical path using spreadsheet methodology.”
“Automatic recognition of elbow musculoskeletal disorders using cloud [38] T. Mahmud, D. Islam, M. Begum, S. Das, L. Dey, and K. Barua, “A
application,” in Proceedings of the 8th International Conference on decision concept to support house hunting,” International Journal of
Software and Information Engineering, 2019, pp. 80–84. Advanced Computer Science and Applications(IJACSA), vol. 13, no. 10,
[19] Y. L. Thian, Y. Li, P. Jagmohan, D. Sia, V. E. Y. Chan, and R. T. pp. 768–774, 2022.
Tan, “Convolutional neural networks for automated fracture detection [39] M. J. A. Patwary, S. Akter, and T. Mahmud, “An expert system to
and localization on wrist radiographs,” Radiology: Artificial Intelligence, detect uterine cancer under uncertainty,” IOSR Journal of Computer
vol. 1, no. 1, p. e180001, 2019. Engineering (IOSR-JCE), e-ISSN, pp. 2278–0661, 2014.
[20] N. Pradhan, V. S. Dhaka, and H. Chaudhary, “Classification of human [40] T. Mahmud, J. Sikder, and S. R. Naher, “Decision support system for
bones using deep convolutional neural network,” in IOP conference house hunting: A case study in chittagong,” in Proceedings of the Future
series: materials science and engineering, vol. 594, no. 1. IOP Technologies Conference (FTC) 2020, Volume 2. Springer, 2021, pp.
Publishing, 2019, p. 012024. 676–688.
[21] Y. Shao and X. Wang, “A two stage method for abnormality diagnosis [41] T. Mahmud, I. Ara, R. Chakma, K. Barua, D. Islam, M. S. Hossain,
of musculoskeletal radiographs,” in International Conference on Pattern A. Barua, and K. Andersson, “Design and implementation of an ul-
Recognition and Artificial Intelligence. Springer, 2020, pp. 610–621. trasonic sensor-based obstacle avoidance system for arduino robots,”
[22] H. El-Saadawy, M. Tantawi, H. A. Shedeed, and M. F. Tolba, “A hybrid in 2023 International Conference on Information and Communication
two-stage gng–modified vgg method for bone x-rays classification and Technology for Sustainable Development (ICICT4SD). IEEE, 2023, pp.
abnormality detection,” IEEE Access, vol. 9, pp. 76 649–76 661, 2021. 264–268.
[23] A. Sezer and H. B. Sezer, “Convolutional neural network based diagnosis
of bone pathologies of proximal humerus,” Neurocomputing, vol. 392,
pp. 124–131, 2020.