Utilizing InceptionV3 for Categorizing Cervical Spine Fractures and

Assessing Accuracy Against a Convolutional Neural Network

Kaviya V. H and P. V. Parimala

Saveetha Institute of Medical and Technical Sciences, Saveetha University, Chennai, Tamil Nadu, 602105, India

Keywords: Cervical Spine Fracture, Convolutional Neural Network, Deep Learning, Health, Novel InceptionV3, CT.

Abstract: The intent of this study is to compare the accuracy of Novel InceptionV3 and Convolutional Neural Networks

in detecting cervical spine fractures from CT images. The two groups of learning models proposed in this

study are the Novel InceptionV3 deep learning model and the Convolutional Neural Network (CNN). Cervical

fracture is the dataset taken for the analysis which is obtained from the open source Kaggle repository with a

sample size of 4200 CT images. In which 3800 images were given to train the model and 400 images to

evaluate the model. With the value of G power = 0.8 with 95% confidence interval the experiment is iterated

tenfold. The classification accuracy yielded by the proposed algorithm Novel InceptionV3 is 94.56% while

CNN obtained an accuracy of 77.32%. The T-test (p<0.001, two tailed) shows that Novel InceptionV3 appears

to have more significance than CNN. Conclusion: The study investigated the performance of two deep

learning models in predicting cervical spine fractures with higher accuracy. The outcome indicates that Novel

InceptionV3 is more effective in comparison with convolutional neural networks.

1 INTRODUCTION

Cervical fractures if untreated can result in lifelong

paralysis, which is highly dangerous and potentially

fatal (Campagnolo et al. 2011). In the US, there are

an estimated 12,000 new cases of cervical spine

injuries each year, 42% of which are caused by

automobile accidents. Of these cervical spine injuries,

sports account for roughly 8% (Belval 2015). In

North America, injuries to the cervical spine result in

more than a million visits to emergency departments

annually (Milby et al. 2008). The initial step in

treating them is finding the fracture (Dambhare and

Kumar 2022). Therefore, an automated cervical spine

fracture detection system is highly important for early

diagnosis in today’s world. The applications of deep

learning have made incredible progress in the early

diagnosis of Cervical fractures and are

revolutionizing the healthcare industry and enabling

clinicians to treat patients well using clinical data

(Davenport and Kalakota 2019).

Around 1100 papers in Research gate and 600

Science direct articles in over the preceding five years

have been published that are pertinent to the detection

of cervical spine fractures. Various techniques were

used to improve the model's performance. One in

which Hojjat used an approach by equipping deep

sequential learning techniques for identifying

fractures on the cervical part of the spinal column on

CT scans, with a 70.92% accuracy rate (Salehinejad

et al. 2021a). Guillermo et al. proposed two deep

learning models VGG16 and ResNet18 to precisely

predict fractures on sagittal radiographic images. The

accuracies obtained were 88%(ResNet18) and

84%(VGG16) (Rosenberg et al. 2022). A study to

assess the accurate estimation and error rate analysis

of a Deep neural network to discover the presence of

fractures on the cervical spine was performed and

attained an accuracy of 54.9% (Hodler, Kubik-Huch,

and von Schulthess 2020; Voter et al. 2021).The

ResNet 50 and Bidirectional Long Short-Term

Memory (BLSTM) models were combined using an

ensemble methodology by Hojjat and others,

demonstrating the effectiveness of deep neural

networking models in tackling this issue (Salehinejad

et al. 2021b). Earlier algorithms were unable to

identify severe fracture locations due to the dataset's

imbalance. The most significant limitations of earlier

studies that reduced the generalizability of the results

were research design and selection bias, hence had an

impact on classification performance. The intent of

this research is to contrast the effectiveness and

functionality of a CNN and a deep learning model

called Novel InceptionV3 using CT to ascertain

which model is preferable at classifying cervical

spine fractures.

H., K. and Parimala, P.

Utilizing InceptionV3 for Categorizing Cervical Spine Fractures and Assessing Accuracy Against a Convolutional Neural Network.

DOI: 10.5220/0012543000003739

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

In Proceedings of the 1st International Conference on Artiﬁcial Intelligence for Internet of Things: Accelerating Innovation in Industry and Consumer Electronics (AI4IoT 2023), pages 583-588

ISBN: 978-989-758-661-3

583

Figure 1: Displays the model’s flowchart performed in

Novel InceptionV3 and Convolutional Network.

2 METHODS & MATERIALS

The comparative analysis was performed in the

Programming Laboratory of Saveetha School of

Engineering, Saveetha Institute of medical and

technical Sciences (SIMATS). Novel InceptionV3

and Convolutional Neural Network are the two neural

networks taken for the analysis with size of (N=10)

samples. Using the aforementioned techniques, the

experiment was iterated over ten times. The G-power

was established to be 0.8 with 95% confidence

interval for the supplied data samples.

The HP Pavilion Laptop 14ec with Windows 11

Home had an AMD Ryzen 5 series processor, 8GB

DDR4-3200 MHz RAM, a 64-bit OS, x64 based

processor and AMD Radeon Graphics served as the

procedure's infrastructure. Google Colaboratory

served as the platform to train the model. Adam is the

optimizer which was used to compile the model.

Table 1: Accuracy values for the groups: Inception V3 and

CNN model.

S.NO ACCURACY

Inception V3 CNN

1 74.95 72.08

2 92.80 73.3

3 94.38 77.50

4 96.09 74.17

5 96.17 76.25

6 96.81 79.58

7 97.59 77.08

8 97.59 82.92

9 97.24 79.58

10 97.97 80.83

Average 94.56 77.32

Table 2: Values obtained from the performance metrics on

evaluating the models.

S.NO METRICS Ince

tionV3 CNN

1 Accurac

94.5% 83.2%

2 Precision 100% 83%

4 F1-score 94.7% 83.2%

5 Sensitivit

90% 83.4%

6 S

ecificit

100% 83.08%

The Cervical fracture dataset used for the analysis

was obtained from kaggle suggested by V3 and CNN

model. (Sairam 2022) and consists of 4,200 cervical

spine CT images sized 224x224. They were

subdivided into fracture and normal samples. The

model was trained on 80% of the image samples and

evaluated on 20% of the image samples. The CT

images of the train and test were then categorized into

normal and fracture samples.

2.1 Inception

InceptionV3 is a transfer learning model for image

analysis. This network is an improved version of the

InceptionV1

model. There are 48 layers in total as

shown in Fig. 1. It is more efficient, has deeper

networks than the Inception V1 and V2 models, but

its speed is not compromised. It is less expensive in

terms of computation. It consists of Convolutional

layers which are factored, smaller, and asymmetric to

lower the computational efficiency. During training,

an auxiliary classifier serves as a regularizer between

layers, and the loss it incurs is added to the primary

network loss. Feature maps are concatenated in

parallel with a stride two convolution layer and a

max-pooling layer.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

584

Figure 2: Architectural representation of the model InceptionV3.

2.2 Convolutional Neural Network

Convolutional Neural Network, or CNN has three

layers: convolutional, pooling, and fully connected

(FC).

Convolutional is the first layer to extract features

and FC is the second which ensures every input of the

input vector influences every output of the output

vector. From the CL to the FC layer, CNN increases

with complexity. With increasing complexity, the

CNN can capture more intricate, larger portions of an

image till discovering the complete object.

2.3 Statistical Analysis

The comparison was conducted using Statistical

Package for the Social Sciences software 26. This tool

was designed to undertake statistical analysis for the

given data. In its stage of development, it offers a vast

library of AI statistics, open-source scalability, and

the ability to evaluate the mean accuracy of different

algorithms (SPSS Software, n.d.). Adequacy is the

dependent variable, and the accuracy of Novel

InceptionV3 is the independent variable. The

comparison of the two independent groups

InceptionV3 and Convolutional Neural Network was

done using the Independent Samples T-test to see if

there is a proof that the means of the corresponding

populations are significantly dissimilar.

Figure 3a demonstrates the accuracy of the model

trained and validated of both the models using a line

graph proving the accuracy of InceptionV3 (94.56%)

is higher compared to CNN (77.32%). The training

and validation losses of InceptionV3 and CNN are

compared in Fig. 3b. From the confusion matrix of

InceptionV3 in Fig. 4a &b and CNN, it can be seen

that the values of True.

Figure 1: Flowchart Inception V3 and CNN.

Positive and True Negatives are greater in

InceptionV3 in contrast to CNN inferring that

InceptionV3 has detected the cases more accurately

than CNN. Fig. 4c demonstrates the mean accuracy

and loss of InceptionV3 vs CNN with the group

plotted on the X axis, and Y axis showing the mean

accuracy and loss. Novel InceptionV3 seemed to have

a higher accuracy in detecting cervical fractures using

CT images compared to CNN.

3 DISCUSSION

Novel InceptionV3 obtained an accuracy of 97%

while CNN attained an accuracy of 80%, proving that

Novel InceptionV3 is much more accurate. With

p=0.001, it is noted that Novel InceptionV3

functioned better than expected and was more

effective in detecting cervical spine fractures.

According to research, the prevalence of

undetected fractures in the spine lies from 19.5% -

45%. Pranata and others created two CNN-based

models for calcaneal fracture classification using CT

radiographic images. The included model is a

potential tool for future usage in automated diagnosis

with accuracy of 79%, and 72.9% of specificity

(Pranata et al. 2019). In one study, a computer-aided

technique was suggested for identifying fractures in

calcaneus on Computed Tomography scan images.

They opted for the Sanders fracture classification

system, which makes use of color segmentation to

identify and classify calcaneus fragments. The model

has an accuracy of 86% (Zhang et al. 2018). Some

preliminary studies have demonstrated CNNs are a

suitable tool for fracture prediction on radiographic

Utilizing InceptionV3 for Categorizing Cervical Spine Fractures and Assessing Accuracy Against a Convolutional Neural Network

585

images. (Olczak et al. 2017) used a network trained

on a variety of hand, wrist, and ankle radiographs to

achieve an accuracy of 83% in fracture detection.

(Kim and MacKinnon 2018) achieved an area under

the curve (AUC) of 0.954 with a model trained on

1389 lateral wrist radiographs. According to studies,

the prevalence of undiagnosed spine fractures ranges

from 19.5% to 45%. (Muehlematter et al. 2019) was

using lumbar and thoracic CT images to classify,

detect, and locate vertebral spine fractures, as well as

assess lumbar vertebral bone density. The accuracy of

healthy/unhealthy vertebrae was poor, with an AUC

of 0.5. The sensitivity for compression fracture

identification and localization was 0.957, with a

falsified rate of 0.29 per patient.

Table 3: Independent Sample T-Test is applied for the data set fixing confidence interval as 95% and Significance as p<0.001

(p<0.05) (2-tailed).

Figure 3: (a) Comparison of Training and Validation Accuracies of InceptionV3 and CNN (b) Comparison of Training and

Validation Losses of InceptioNV3 and CNN.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

586

Figure 4: Visualization of the true and predicted cases using confusion matrix of (a) InceptionV3 (b)CNN (c) Bar chart

representing the comparison of Mean accuracy of InceptionV3 & CNN for CT scans.

The amount of data consumed is indeed low when

adopting models from different multi-layer neural

networks, and the test dataset's modest size had an

impact on classification accuracy when evaluating the

model. Future research could be primarily focused on

developing deep learning models that can fit in and

get trained more quickly while utilizing smaller

datasets. Fig. 4c. Bar chart representing the

comparison of Mean accuracy of InceptionV3 and

Convolutional neural network model in cervical spine

fracture detection using CT scans. InceptionV3

appears to produce better results with standard

deviation. X Axis: InceptionV3 vs Convolutional

neural network (CNN) and Y axis: Mean Accuracy of

detection SD = ±2 and confidence interval of 95%.

Table 4: Statistical computation of independent samples

tested among InceptionV3 and CNN deep learning models.

The mean accuracy of InceptionV3 is 94.5 and CNN is

77.329 Standard Deviation of InceptionV3 is 5.68 and CNN

is 4.09. The T-test for comparison for InceptionV3standard

error mean is 1.79 and CNN is 1.29.

Group N Mean

Std.

Deviation

Std. Mean

Erro

Accuracy

InceptionV3 10 94.56 5.687 1.7986

CNN 10 77.329 4.091 1.293

Loss

InceptionV3 10 1.361 1.254 0.3967

CNN 10 4.515 0.862 0.2725

4 CONCLUSIONS

The experimental finding demonstrates that the Novel

InceptionV3 performs better than CNN, with an

accuracy of 94.56% as opposed to the value of

77.32%. In terms of workflow effectiveness, the

proposed approach has a wide range of potential

applications in medical imaging. The analysis

exhibits the applicability of deep learning models that

can handle the issue of large datasets with improvised

accuracy.

REFERENCES

Belval, Luke. (2015). “Cervical Spine Injury.” Korey

Stringer Institute. March 4, 2015. https://ksi.ucon

n.edu/emergency-conditions/cervical-spine-injury/.

Campagnolo, Denise I., Steven Kirshblum, Mark S. Nash,

Robert F. Heary, and Peter H. Gorman. (2011). Spinal

Cord Medicine. Lippincott Williams & Wilkins.

Dambhare, Shruti, and Sanjay Kumar. (2022). “Machine

Learning in Healthcare.” Machine Learning and Deep

Learning in Efficacy Improvement of Healthcare

Systems. https://doi.org/10.1201/9781003189053-1.

Davenport, Thomas, and Ravi Kalakota. (2019). “The

Potential for Artificial Intelligence in Healthcare.”

Future Healthcare Journal 6 (2): 94–98.

G. Ramkumar, R. Thandaiah Prabu, Ngangbam Phalguni

Singh, U. Maheswaran, (2021), Experimental analysis

of brain tumor detection system using Machine learning

approach, Materials Today: Proceedings,ISSN 2214-

7853,https://doi.org/10.1016/j.matpr.2021.01.246.

Utilizing InceptionV3 for Categorizing Cervical Spine Fractures and Assessing Accuracy Against a Convolutional Neural Network

587

Hodler, Juerg, Rahel A. Kubik-Huch, and Gustav K. von

Schulthess. (2020). Diseases of the Brain, Head and

Neck, Spine 2020–2023: Diagnostic Imaging. Springer

Nature.

Kim, D. H., and T. MacKinnon. (2018). “Artificial

Intelligence in Fracture Detection: Transfer Learning

from Deep Convolutional Neural Networks.” Clinical

Radiology 73 (5): 439–45.

Kumar M, M., Sivakumar, V. L., Devi V, S.,

Nagabhooshanam, N., &Thanappan, S. (2022).

Investigation on Durability Behavior of Fiber

Reinforced Concrete with Steel Slag/Bacteria beneath

Diverse Exposure Conditions. Advances in Materials

Science and Engineering, 2022.

Milby, Andrew H., Casey H. Halpern, Wensheng Guo, and

Sherman C. Stein. (2008). “Prevalence of Cervical

Spinal Injury in Trauma.” Neurosurgical Focus 25 (5):

E10.

Muehlematter, Urs J., Manoj Mannil, Anton S. Becker,

Kerstin N. Vokinger, Tim Finkenstaedt, Georg

Osterhoff, Michael A. Fischer, and Roman

Guggenberger. (2019). “Vertebral Body Insufficiency

Fractures: Detection of Vertebrae at Risk on Standard

CT Images Using Texture Analysis and Machine

Learning.” European Radiology. https://doi.org/10.

1007/s00330-018-5846-8.

Olczak, Jakub, Niklas Fahlberg, Atsuto Maki, Ali Sharif

Razavian, Anthony Jilert, André Stark, Olof

Sköldenberg, and Max Gordon. (2017). “Artificial

Intelligence for Analyzing Orthopedic Trauma

Radiographs.” Acta Orthopaedica 88 (6): 581–86.

Rosenberg, Guillermo Sánchez, Andrea Cina, Giuseppe

Rosario Schiró, Pietro Domenico Giorgi, Boyko

Gueorguiev, Mauro Alini, Peter Varga, Fabio

Galbusera, and Enrico Gallazzi. (2022). “Artificial

Intelligence Accurately Detects Traumatic

Thoracolumbar Fractures on Sagittal Radiographs.”

Medicina 58 (8). https://doi.org/10.3390/medicin

a58080998.

S. G and R. G, "Automated Breast Cancer Classification

based on Modified Deep learning Convolutional Neural

Network following Dual Segmentation," 2022 3rd

International Conference on Electronics and

Sustainable Communication Systems (ICESC),

Coimbatore, India, 2022, pp. 1562-1569, doi:

10.1109/ICESC54411.2022.9885299.

Sairam, Vuppala Adithya. 2022. “Spine Fracture Prediction

from C.T.” https://www.kaggle.com/vuppalaadithya

sairam/spine-fracture-prediction-from-xrays.

Salehinejad, Hojjat, Edward Ho, Hui-Ming Lin, Priscila

Crivellaro, Oleksandra Samorodova, Monica Tafur

Arciniegas, Zamir Merali, et al. (2021)a. “Deep

Sequential Learning For Cervical Spine Fracture

Detection On Computed Tomography Imaging.” 2021

IEEE 18th International Symposium on Biomedical

Imaging (ISBI). https://doi.org/10.1109/isbi48211.

2021.9434126.

Voter, A. F., M. E. Larson, J. W. Garrett, and J-P J. Yu.

(2021). “Diagnostic Accuracy and Failure Mode

Analysis of a Deep Learning Algorithm for the

Detection of Cervical Spine Fractures.” AJNR.

American Journal of Neuroradiology, June.

https://doi.org/10.3174/ajnr.A7179.

Yadollahi, Mahnaz, Shahram Paydar, Haleh Ghaem,

Mohammad Ghorbani, Seyed Mohsen Mousavi, Ali

Taheri Akerdi, Eimen Jalili, et al. (2016).

“Epidemiology of Cervical Spine Fractures.” Trauma

Monthly 21 (3): e33608.

Zhang, Zhirui, Shujie Liu, Mu Li, Ming Zhou, and Enhong

Chen. (2018). “Bidirectional Generative Adversarial

Networks for Neural Machine Translation.”

Proceedings of the 22nd Conference on Computational

Natural Language Learning. https://doi.org/10.

18653/v1/k18-1019.

AI4IoT 2023 - First International Conference on Artiﬁcial Intelligence for Internet of things (AI4IOT): Accelerating Innovation in Industry

and Consumer Electronics

588