Volume 8, Issue 10, October – 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Develop an Extended Model of CNN

Algorithm in Deep Learning for Bone
Tumor Detection and its Application
(Area of Focus: Deep Learning)

1
Elyse MUGABO; Dr. Wilson MUSONI (PhD)2
Masters of Science with honors in Information Technology, at University of Kigali, Rwanda

Abstract:- Deep leaning in orthopedic surgery has gained The use of convolutional neural networks (CNNs) has led to
mass interest over the last decade or so. In prior studies, highly accurate and efficient classification of different types
researchers have demonstrated that deep learning in of tumors based on medical images such as X-rays, CT scans,
orthopedics can be used for different applications such as and MRI. As bone tumors can be difficult to distinguish from
fracture detection, bone tumor diagnosis, detecting hip normal bone tissue, the accuracy and efficiency provided by
implant mechanical loosening, and grading osteoarthritis. deep learning algorithms can significantly improve early
As time goes on, the utility of deep learning algorithms, diagnosis and treatment of bone tumors. Overall, deep
continues to grow and expand in orthopedic surgery. The learning has the potential to transform the field of medical
purpose of this research was to develop an extended imaging analysis and improve patient outcomes.
model of CNN algorithm in deep learning for bone tumor
detection and its application. Bone tumors can be II. METHODOLOGY
malignant growths. Despite the fact that it can happen in
any bone, it frequently happens in long bones like the  Data Collection Methods and Instruments/ Tools
arms and legs. Although the exact source of this The practice of gathering information using specified
malignant tumor is yet unknown, doctors believe that procedures in order to react to the study's predetermined
DNA abnormalities within the bones are to blame. In research subject is known as data collecting. In this study, the
addition to destroying good bodily tissue, this results in researcher used mixed method (both qualitative and
immature, crooked, and diseased bone. When a bone quantitative) and examined secondary data.
tumor is suspected, the first test is a bone X-ray. The
greatest method for detecting cancer in the bones is  Data Analysis
through imaging and X-ray scans. The recommended The process of discovering solutions through
procedure that can provide a certain diagnosis is a investigation and interpretation is known as data analysis.
biopsy. This labor-intensive and challenging process can Understanding survey and administrative source results and
be mechanized. We presented a number of supervised presenting data information require data analysis. Data
deep learning techniques and chose the suitable model. analysis is anticipated to provide light on the subject of the
To find bone cancer, a selection is made using the study and the respondents' perceptions, as well as to increase
weighted average of user data. Using the residual neural readers' understanding of the subject and pique their interest
network (ResNet101) technique, we extended the models in this portion of the research. Jupyter notebook was used to
that were chosen and they met the expectations with the analyze the data and present the results by measuring the
maximum accuracy (90.36%) and precision (89.51%), performance metrics.
respectively, for the prediction tasks.
 Research Design
Keywords:- CNN: Convolutional Neural Networks, ANN: A research design is a plan or blueprint which shows
Artificial Neural Networks, MRI: Magnetic Resonance how data required for the solution of the problem that the
Imaging, AI: Artificial Intelligence. researcher focused on, the procedure and methods for data
collection and analysis, answered the research questions. In
I. INTRODUCTION these lines the research herein the present study, researcher
would employ a combination of descriptive and correlation
Certainly, deep learning has emerged as a promising research design to describe the characteristics of a population
technology for medical imaging analysis, including the under investigation and carefully examine the use CCN
detection and diagnosis of bone tumors. Traditional machine algorithm in bone tumor detection (Hardt,2016).
learning methods rely on expert input to extract features
relevant to identifying bone tumors in medical images, but  Objective:
deep learning algorithms can learn these features on their The objective of this study is to develop an extended
own from raw data. This is especially useful in medical model of a Convolutional Neural Network (CNN) algorithm
imaging where there is a vast amount of data to be analyzed.

IJISRT23OCT295 www.ijisrt.com 661

Volume 8, Issue 10, October – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
for bone tumor detection. Additionally, the study aims to  Model Evaluation:
explore the practical application of the developed model in Evaluate the trained model using the validation dataset.
real-world scenarios. Monitor performance metrics such as accuracy, precision,
recall, F1 score, and area under the ROC curve. Conduct
 Data Collection: multiple epochs of training and validation to fine-tune the
Gather a comprehensive dataset of medical images that model's hyperparameters and ensure optimal performance.
includes various types of bone tumors and normal bone
images. Ensure that the dataset represents a diverse range of  Testing and Application:
cases, including different tumor types, sizes, and locations. Test the final trained model using the independent
Collect the data from medical archives, hospitals, or curated testing dataset to assess its performance on unseen data.
datasets. Evaluate metrics like accuracy, sensitivity, specificity, and
area under the ROC curve on the test dataset. Additionally,
 Data Preprocessing: deploy the model in real-world scenarios, such as using it in a
Preprocess the collected medical images to enhance software application or a medical imaging system, to
their quality and enable compatibility with deep learning demonstrate its practical application.
algorithms. Apply techniques such as resizing, cropping,
normalization, and noise removal to ensure consistent image  Ethical Considerations:
quality throughout the dataset. Address ethical considerations such as data privacy,
patient consent, and responsible use of AI in healthcare.
 Splitting the Dataset: Ensure compliance with ethical guidelines and regulations
Divide the dataset into training, validation, and testing throughout data collection, model deployment, and
subsets. Use an appropriate split, such as 80% for training, dissemination of research findings.
10% for validation, and 10% for testing. This split allows
training the model, optimizing hyper-parameters, and  CNN (Convolutional neural networks)
evaluating the performance effectively. CNN are a specific type of neural network designed for
processing data that possesses a grid-like structure. Examples
 Model Architecture: of such data include time series data, which is regarded as a
Design an extended CNN architecture that builds upon 1D grid with regularly spaced samples over time, and image
existing CNN models for bone tumor detection. Consider data, which is treated as a 2D grid composed of pixels. CNNs
incorporating additional layers, such as residual connections, have demonstrated remarkable effectiveness in practical
attention mechanisms, or novel convolutional filters, to applications. The term "convolutional neural network" refers
enhance the model's performance. Experiment with different to the network's utilization of the mathematical operation
architecture modifications to find the best approach. called convolution. Convolution belongs to the broader
category of linear operations. Convolutional networks can be
 Training the Model: viewed as neural networks that incorporate at least one layer
Train the extended CNN model using the training where convolution operations are employed instead of
dataset. Utilize loss functions like categorical cross-entropy general matrix multiplication (LeCun, 1989).
or focal loss, and optimizer algorithms such as Adam or
RMSprop. Implement appropriate techniques like data  Methodology for Bone Tumor Detection using
augmentation to augment the training dataset and increase its Convolutional Neural Network Models
diversity.

IJISRT23OCT295 www.ijisrt.com 662

Volume 8, Issue 10, October – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig 1 Methodology for Bone Tumor Detection using CNN Models

 VGG16 (Visual Geometry Group 16) everyone who is impacted by their study or the results report.
The VGG-16 model is a convolutional neural network Several ethical rules are typically applied to researchers.
(CNN) consisting of 16 layers. It is widely recognized as one
of the top-performing and efficient models available today. In  Data Privacy and Security:
contrast to architectures with different params, VGG16 Ensure that patient privacy is protected throughout the
model relies on Convolutional Network layers using a 3x3 data collection, handling, and storage process. Follow data
size of the kernel. This model is particularly valuable because protection regulations and guidelines, such as anonymizing
it can be found and downloaded online for various systems or de-identifying patient information, encrypting sensitive
and applications. Its simplicity stands out compared to other data, and implementing secure data storage practices.
comprehensive models that have been developed.
 Informed Consent:
For the VGG-16 model, the minimum input image size Obtain informed consent from patients or individuals
required is 224x224 pixels with three channels. In neural whose medical images are included in the dataset. Ensure
networks, optimization techniques are employed to assess the that participants understand the purpose of the study, how
involvement of a neuron by calculating the weighted sum of their data will be used, potential ramifications, and their
its inputs. A kernel function is utilized to introduce non- rights regarding data handling and privacy.
linearity in the output neuron. Neurons in a neural network
interact with weights, biases, and training techniques. The  Bias and Fairness:
connection weights between neurons are adjusted based on Be aware of potential bias during data collection and
the desired output. Input images and activation functions play model development stages. Take measures to address and
a significant role in introducing nonlinearity into artificial mitigate biases in the dataset, such as ensuring diverse
neural networks. representation of demographics and accounting for potential
disparities in disease prevalence across different populations.
 Ethical Considerations Regularly evaluate the model's performance across various
The data collected for this study is protected with the subgroups to ensure fairness and avoid discrimination.
help of ethical consideration. Research ethics is the
obligation of researchers to be truthful and considerate to

IJISRT23OCT295 www.ijisrt.com 663

Volume 8, Issue 10, October – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Transparency and Explainability: and always uses the same padding and maxpool layer of 2
Strive for transparency in the development and × 2 filter with stride 2. It is 16-layer Convolutional Neural
implementation of the deep learning model. Clearly Network architecture; the “16” in VGG16 denotes that it
document the methodology, architecture, and any includes 16 layers with weights.
modifications made to the CNN algorithm. Consider  VGG19: is a Convolutional Neural Network architecture
incorporating interpretability techniques to provide which is widely used for image classification and other
explanations for the model's decisions, enabling healthcare applications that has been trained on the Image dataset It
professionals to understand and trust the model's output. was created at Oxford University by the Visual Geometry
Group and is distinguished by the use of extremely tiny (3
 Ethical use of AI: × 3) and deep (3 × 3) convolutional filters (19 layers).
Ensure that the developed model is used ethically and The architecture is straightforward yet effective and has
responsibly. Consider potential risks, limitations, and served as the foundation for several cutting-edge models.
implications of the model's application in real-world  DenseNet201: A convolutional neural network
scenarios. Monitor and address any unintended consequences architecture called DenseNet201 is a development on the
that may arise from using the model, and regularly update DenseNet framework. It was created in 2017 by Gao
and improve the model as needed. Huang and colleagues. The primary concept of DenseNet
is to feed-forward link each layer to every other layer. As
 Continual Evaluation and Validation: a result, a dense connectivity network is produced, which
Regularly assess and validate the performance of the enhances feature reuse and lessens the issue of
extended CNN model, both during development and after disappearing gradients. A DenseNet architecture variant
deployment. Maintain ongoing monitoring of the model's called DenseNet201 contains 201 layers. It is frequently
accuracy, reliability, and generalizability. Address any used in computer vision applications like object
performance issues, biases, or unintended consequences that identification and picture categorization.
may emerge.  ResNet 101: Convolutional Neural Network architecture,
in 2015. The ResNet model, which stands for “Residual
 Collaboration and Collaboration with Medical Network”, is a version of this one. The introduction of
Professionals: “residual connections”, which enable the network to study
Collaborate closely with medical professionals, residual functions with relevance to the layer inputs
radiologists, and other healthcare practitioners throughout the instead of learning unsourced functions, is the main
research and development process. Seek their expertise, novelty of ResNet. As a result, training very deep
guidance, and feedback to ensure the model aligns with networks (152 layers or more) is now feasible without
clinical needs, enhances patient care, and adheres to ethical experiencing diminishing gradient issues. A ResNet
standards. variant with 101 layers is known as ResNet101. It has
broken previous records for image classification
 Responsible Reporting and Dissemination: performance on the ImageNet dataset and is frequently
Accurately report and transparently share the findings used for image classification and object recognition
of the research study, including any limitations or challenges applications
encountered during the development of the extended CNN
model. Avoid sensationalism or overstatements about the  Data Analysis and Presentation of Findings
model's capabilities. Share the research findings within the Performance metrics were used as a data analysis
scientific community, enabling peer review, replication, and technique in this study, and the produced graphs were
further advancements in the field. examined. Data were loaded and split with split sizes of 0.8
and 0.2 into training and testing groups. The GPU Tesla T4
These ethical considerations are crucial to ensure with an image size of [256,256] was used, and the batch size
patient safety, privacy, and fairness while promoting was set at 64. 'VGG16', 'VGG19', 'DenseNet201', and
responsible development of an extended model of CNN 'ResNet101' ImageNet models were used for classification,
algorithm in deep learning for bone tumor detection and its with the output layer set to softmax activation function. 'Non-
application in the healthcare domain. Tumor,' 'Non-Viable-Tumor,' and 'Viable-Tumor' were the
three groups into which the photos in the dataset were split.
 Applied Algorithms to Extend CNN Models Adam optimizer and tf.keras.losses were employed. AS a loss
function, categorical Cross-entropy was used.
 VGG16: Convolutional Neural Network (CNN)
architecture Visual Geometry Group-16 concentrates on
having convolution layers of 3 × 3 filter with a stride 1

IJISRT23OCT295 www.ijisrt.com 664

Volume 8, Issue 10, October – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
III. DATA PRESENTATION

 Validation Accuracy Curve for all Models

The figure below represented the accuracy achieved by all the models, acceptable and shows the Validation Accuracy of all
Models ‘VGG16', 'VGG19', 'DenseNet201', and 'ResNet101'.

Fig 2 Validation Accuracy Curve for all Models

Fig 3 Accuracy of all Models

Fig 4 Validation Loss curve for all models

IJISRT23OCT295 www.ijisrt.com 665

Volume 8, Issue 10, October – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Comparative Interpretation of Findings between VGG16,VGG19,DenseNet201 and ResNet101 Models

Table 1 Performance of Algorithms Performed on the Dataset

Parameters VGG16 VGG19 DenseNet201 ResNet101
Accuracy 86.69% 89.44% 87.15% 90.36%
F1 85.50% 88.33% 85.50% 89.35%
Precision 84.31% 87.79% 84.38% 89.51%
Recall 87.21% 88.94% 87.15% 89.59%
AUC 0.909 0.931 0.909 0.946
Vscore 2.361 2.506 1.970 2.720
Comments Here VGG16 ResNet101 is the best
in the result Predict the bone tumor
About the accuracy Compare to other
on the highest models

Fig 5 Evaluation of Algorithms Performed on the Dataset

 Confusion Matrix of Models

To find misclassifications, compares a classification
model's predicted and actual values. The confusion matrix
helped me in the performance optimization of my models.
The figures below show the visualization of the performance
on the models. TensorFlow was utilized to build these
classification models. Fig 7 Confusion Matrix for VGG19

Fig 6 Confusion Matrix for VGG16 Fig 8 Confusion Matrix for DenseNet201

IJISRT23OCT295 www.ijisrt.com 666

 Volume 8, Issue 10, October – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
ResNet101 Convolutional Neural Network architecture was
better suited for the detection bone tumor

The DenseNet201 model has an accuracy of 87.15%, an

F1-score of 85.5%, an AUC of 0.9093, a precision of
84.38%, a recall of 87.15%, and vscore of 1.907 which is the
lowest amongst the tested models, the metrics are similar as
VGG16 but the vscore is low because of time factor. The
ResNet101 model has an accuracy of 90.36%, an F1-score of
89.35%, an AUC of 0.9461, a precision of 89.51%, a recall of
89.59%, and vscore of 2.72 which is the highest as compared
to other models, the residual neural network (ResNet101)
fared well as compared to other models and with lowest time
take to train the model and highest accuracy.

Notably, it is essential to emphasize that the ResNet101

model outperformed the CNN model in terms of accuracy
and F1 score. This implies that ResNet101 model delivered
more precise results and achieved a better balance between
precision and recall in the specific context of bone tumor
detection.

Fig 9 Confusion Matrix for DenseNet101 In conclusion, after meticulously analyzing performance
metrics, plotted graphs, and confusion matrices, I strongly
 Model Evaluation recommend that Rwanda Military Hospital to deploy the
ResNet101 Approach model for bone tumor detection. The
ResNet101 model demonstrated exceptional accuracy and
precision on the test set. By deploying this model, the
hospital can ensure reliable and accurate bone tumor
detection outcomes.

REFERENCES

[1]. AlDubayan SH, Conway JR, Camp SY, Witkowski L,

Kofman E, Reardon B, et al. (2020). Detection of
Pathogenic Variants With Germline Genetic Testing
Using Deep Learning vs Standard Methods in Patients
With Prostate Cancer and Melanoma. Jama
[2]. Ardila D, Kiraly AP, Bharadwaj S, Choi B, Reicher
JJ, Peng L, et al. (2019) End-ToEnd Lung Cancer
Screening with Three-Dimensional Deep Learning on
Low-Dose Chest Computed Tomography. Nat Med
[3]. Cheng CT, Ho TY, Lee TY, Chang CC, Chou CC,
Chen CC, et al. (2019). Application of a Deep
Learning Algorithm for Detection and Visualization
Fig 10 Comparative of V-Score for all Models of Hip Fractures on Plain Pelvic Radiographs. Eur
Radiol
IV. CONCLUSIONS AND DISCUSSIONS ON THE [4]. Chung SW, Han SS, Lee JW, Oh KS, Kim NR, Yoon
JP, et al. (2018). Automated Detection and
FINDINGS
Classification of the Proximal Humerus Fracture by
For the formation of model, the proposed framework is Using Deep Learning Algorithm. Acta Orthop.
composed to use supervised deep learning methods. And, for [5]. Ferguson JL, Turner SP. (2018). Bone Cancer:
classification problem VGG16, VGG19, DenseNet201 and Diagnosis and Treatment Principles. Am Fam.
ResNet101 algorithms were applied. The VGG16 model has Physician
an accuracy of 86.69%, an F1-score of 85.5%, an AUC of [6]. Ferguson JL, Turner SP. Bone Cancer: Diagnosis and
0.909, a precision of 84.31%, a recall of 87.21%, and vscore Treatment Principles. Am Fam. Physician (2018)
of 2.36 which fared average as compared to other models. [7]. 7.Gulshan V, Peng L, Coram M, Stumpe MC, Wu D,
The VGG19 model has an accuracy of 88.94%, an F1-score Narayanaswamy A, et al. (2016) Development and
of 88.33%, an AUC of 0.9314, a precision of 87.79%, a recall Validation of a Deep Learning Algorithm for
of 88.94%, and vscore of 2.506 which is the second-best Detection of Diabetic Retinopathy in Retinal Fundus
score as compared to other models, here we can see that Photographs. Jama

IJISRT23OCT295 www.ijisrt.com 667

Volume 8, Issue 10, October – 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[8]. He J, L. L. (2021). Functional Interplay Between [23]. Schelb P, Kohl S, Radtke JP, Wiesenfarth M,
Long non-Coding RNAs and Want Signaling Cascade Kickingereder P, Bickelhaupt S, et al. (2019) .
in Osteosarcoma . Cancer Cell Int. Classification of Cancer at Prostate MRI: Deep
[9]. Hong L, Cheng X, Zheng D. (2021) Application of Learning Versus Clinical PI-RADS Assessment.
Artificial Intelligence in Emergency Nursing of Radiology
Patients with Chronic Obstructive Pulmonary Disease. [24]. Sheng G, G. Y. (2021). Osteosarcoma and Matastasis.
Contrast Media Mol Imaging Front Oncol.
[10]. Hong L, Cheng X, Zheng D. (2021) Application of [25]. Tomita N, Cheung YY, Hassanpour S. (2018). Deep
Artificial Intelligence in Emergency Nursing of Neural Networks for Automatic Detection of
Patients with Chronic Obstructive Pulmonary Disease. Osteoporotic Vertebral Fractures on CT scans.
Contrast Media Mol Imaging Comput Biol Med
[11]. Kermany DS, Goldbaum M, Cai W, Valentim CCS, [26]. Tu C, H. J. (2020). Emerging Landscape of Circular
Liang H, Baxter SL, et al. (2018) Identifying Medical RNAS as Biomarkers and Pivotal Regulations in
Diagnoses and Treatable Diseases by Image-Based Osteosarcoma. J Cell Physiol.
Deep Learning. Cell [27]. Wang S, Shi J, Ye Z, Dong D, Yu D, Zhou M, et al.
[12]. Kido S, Hirano Y, Mabu S. (2020). Deep Learning for (2019). Predicting EGFR Mutation Status in Lung
Pulmonary Image Analysis: Classification, Detection, Adenocarcinoma on Computed Tomography Image
and Segmentation. Adv Exp Med Biol Using Deep Learning. Eur Respir J
[13]. Kim DH, MacKinnon T. (2018). Artificial [28]. Wu TT, Zheng RF, Lin ZZ, Gong HR, Li H. (2021) A
Intelligence in Fracture Detection: Transfer Learning Machine Learning Model to Predict Critical Care
from Deep Convolutional Neural Networks. Clin Outcomes in Patient with Chest Pain Visiting the
Radiol Emergency Department. BMC Emerg Med
[14]. Leite TC, Watters RJ, Weiss KR, Intini G. (2021). [29]. Wu TT, Zheng RF, Lin ZZ, Gong HR, Li H. (2021).
Avenues of Research in Dietary Interventions to A Machine Learning Model to Predict Critical Care
Target Tumor Metabolism in Osteosarcoma. J Transl Outcomes in Patient with Chest Pain Visiting the
Med Emergency Department. BMC Emerg Med
[15]. Leite TC, Watters RJ, Weiss KR, Intini G. (2021).
Avenues of Research in Dietary Interventions to
Target Tumor Metabolism in Osteosarcoma. J Transl
Med.
[16]. 16. Li W, Ding Z, Wang D, Li C, Pan Y, Zhao
Y, et al. (2021) Ten-Gene Signature Reveals the
Significance of Clinical Prognosis and Immuno-
Correlation of Osteosarcoma and Study on Novel
Skeleton Inhibitors Regarding MMP9. Cancer Cell Int
[17]. Li W, Ding Z, Wang D, Li C, Pan Y, Zhao Y, et al.
(2021). Ten-Gene Signature Reveals the Significance
of Clinical Prognosis and Immuno-Correlation of
Osteosarcoma and Study on Novel Skeleton Inhibitors
Regarding MMP9. Cancer Cell Int
[18]. Munir K, Elahi H, Ayub A, Frezza F, Rizzi A. (2019).
Cancer Diagnosis Using Deep Learning: A
Bibliographic Review. Cancers (Basel)
[19]. Ogrinc N, Caux PD, Robin YM, Bouchaert E, Fatou
B, Ziskind M, et al. (2021) Direct Water-Assisted
Laser Desorption/Ionization Mass Spectrometry
Lipidomic Analysis and Classification of Formalin-
Fixed Paraffin Embedded Sarcoma Tissues Without
Dewaxing. Clin Chem
[20]. Ogrinc N, Caux PD, Robin YM, Bouchaert E, Fatou
B, Ziskind M, et al.(2021)Direct Water-Assisted Laser
Desorption/Ionization Mass Spectrometry Lipidomic
Analysis and Classification of Formalin-Fixed
Paraffin Embedded Sarcoma Tissues Without
Dewaxing. Clin Chem
[21]. Ottaviani G, J. N. (2009). The Epidemology of
Osteosarcoma . Cancer Treat Res.
[22]. Raman R, Srinivasan S, Virmani S, Sivaprasad S, Rao
C, Rajalakshmi R. (2019). Fundus Photograph-Based
Deep Learning Algorithms in Detecting Diabetic
Retinopathy. Eye (Lond)

IJISRT23OCT295 www.ijisrt.com 668