2022 IEEE 2nd International Conference on Mobile Networks and Wireless Communications (ICMNWC)

Cervical Spine Fracture Detection Using Pytorch

1st M. Bhargavi Sri Bhavya 2nd M. Vani Pujitha 3rd G. Lakshmi Supraja
2022 IEEE 2nd International Conference on Mobile Networks and Wireless Communications (ICMNWC) | 978-1-6654-9111-2/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICMNWC56175.2022.10031629

Computer Science and Engineering Computer Science and Engineering Computer Science and Engineering
Velagapudi Ramakrishna Siddhartha Velagapudi Ramakrishna Siddhartha Velagapudi Ramakrishna Siddhartha
Engineering College Engineering College Engineering College
Vijayawada, India Vijayawada, India Vijayawada, India
mbsbhavya@gmail.com pujitha@vrsiddhartha.ac.in glsupraja20@gmail.com

Abstract—Cervical spine fractures are one of the severe damages to the cervical spine if the necessary care was not
varieties and most typical spinal cord injury. The cervical spine taken in time [1]. Sometimes there are some unstable patients
is fragile and well-assembled structure of stacked bones, who need the immediate treatment from the accident and need
ligaments, nerves and muscles. It consists of seven stacked the immediate observations for not to further delay the
vertebrae labelled as C1 to C7 settling exactly below the skull. operation. Faster and easy identification of any fracture or
The injury to any of the vertebrae can result in tetraplegia or injury in the cervical spine can help in faster treatment to the
quadriplegia which means the absent feeling or movement unstable patients [2].
below the neck. Trauma is the prevalent cause of injury as a
result of motor vehicle accidents, falls, sports related such as From the below Fig.1[3] it is observed that the c-spine is
diving injuries. Non traumatic causes composed of arthritis, upper part of the spinal cord which is connected to the brain
cancer, osteoporosis comprehending compression fractures and nerves. So, any fracture or injury to the cervical spine can
inflammations of spinal cord. Computer Tomography (CT) has affect the brain stem, people may have the problem sensing
been emerging as best quality of care imaging diagnosis vibration and temperature and also can have weak muscles
approach especially evaluating the c-spine. Deep convolution and impaired vision. This goal of the model is to create
neural network is built to extricate the radiological factors from machine learning models that can diagnose and localize
each slice of the CT scan. The features that are extracted down fractures to the seven vertebrae that make up the cervical spine
the CT scans are processed and the fractures are identified with
better than radiologist.
the generated probabilities from the prediction. The model was
trained and evaluated with 2019 CT scans each consists of more
than 200 2D image slices. The outcome of the model suit with the
observations and performance of radiologist on the test set in
the physical world process of diagnosis.

Keywords—Cervical Spinal Injury, Vertebrae, Computer

Tomography, Convolution Neural Network, Medical Image
Diagnosis, DICOM images.

I. INTRODUCTION
Spinal cord is the pivotal structure of the human body as it
carries the nerve signal from the brain to body by connecting
them together. These nerve signals are most crucial to help
you feel and move the body parts from the sensation. The
structure of spinal cord incorporated with Cervical, thoracic,
Lumbar, Sacral and Coccyx vertebrae. Here the cervical
vertebrae are directly connected with brain stem which is
responsible for sending and receiving the nerve signals and
any injury to the cervical spine can cause the absence of Fig. 1. Anatomy of Cervical Spine
feeling or movement below the shoulders or neck.
Computer Tomography (CT): It is an imaging diagnosis
CERVICAL SPINE: The cervical part of the spine has method which make use of unique x-ray apparatus to create
seven stacked bones known as vertebrae, categorized C1, C2, precise pictures, or scans, of areas inside the body. It is also
C3, C4, C5, C6, C7. The upper of the c-spine connects to the known as computerized tomography or computerized axial
skull, and the lower part connects to the upper back at about tomography (CAT). The below Fig.2 [4] shows the CT scan
shoulder level. When viewed from the side, the cervical spine of cervical spine.
curves gently forward and then backward to form a lordotic
curve. C-spine injuries comprises of the neck bones, delicate
discs between the vertebrae, joint muscles, ligaments. The
disorders are brought out by injury or malfunction or
deterioration of the structures in the spine from normal wear
and tear, pressure, smoking, aging. The common symptoms of
injury are neck pain. The pain can be in head, jaw, shoulders
as well as numbness and weakness
People often neglect or ignore the neck pains as normal
and common pains due to stress and work pressure. This
ignorance and negligence sometimes can cause major
Fig. 2. CT scan of Cervical Spine

978-1-6654-9111-2/22/$31.00 ©2022 IEEE

Authorized licensed use limited to: Universidad del Valle. Downloaded on October 28,2024 at 18:59:36 UTC from IEEE Xplore. Restrictions apply.
 DICOM: It stands for Digital Imaging and model has a sensitivity of 79%, which is lower than the
Communication in Medicine [5] which have set a standard for radiologist's 93%. Output is therefore evaluated by
medical imaging information management and radiologists after image evaluation.
communication. Now most of the medical image diagnosis are
done through the MRI’s, X-Ray’s, CT scans, DICOM are Hojjat Salehinejad et al. [9] derived deep sequential
mostly used for transmitting the medical images and storing learning for the identification or detection of injuries and
them by allowing the integration with servers, scanners, fractures in the cervical region of spine in computed
workstations, picture archiving communication systems tomography. The authors state that cervical spine fractures
(PACS) which are medical imaging devices. represent a medical emergency that can result in persistent
paralysis and further to death. Faultless diagnosis with
Convolution Neural Networks: It is a Deep Learning surmised fractures by CT is crucial for patient’s health
Algorithm [6] that take an image as an input and assigns management. In this study, the authors derived a deep CNN
learnable biases for different objects in the image which make with continuing storage layers which are bidirectional for
it proficient to differentiate from the other images. In this automatic identification and detection of fractures in axial CT
model deep convolution neural network layers are used for images of CT spine. Validation of the data show classification
extracting the features from the CT scans of the cervical spine accuracies of 70.92% and 79.18%. This approach’s main
and identifying whether there are any fractures that exist in the drawback is the high estimation of false positives, leading to
cervical spine. The layers that are used are conv3d, faulty predictions across cases. If an image in one case is false
maxPool3d, Batchnormal3d, ReLu, Adaptive pool, Linear positive, the whole case leads to false positive. These outputs
layer (output layer). demonstrate the significance of incorporating mundane
features into training. The high performance on unbalanced
PyTorch Framework: It is an Optimized deep learning datasets is mainly because of dataset being biased mostly
tensor library based on Python and Torch, mainly used for towards negative cases and images.
GPU and CPU powered applications. PyTorch [7] is chosen
over other deep learning frameworks such as TensorFlow and Sk. Hasane Ahammad et al. [10] derived a model for a
Keras because it uses a dynamic computational graph and is hybrid CNN-based segmentation and boosting classifier of
completely Pythonic. This allows scientists, developers, and spinal cord injury data from real-time sensors. Spinal cord
neural network debuggers to run and test pieces of code in real disease prediction using convolutional neural networks has
time. So, users don't have to wait until all code is implemented emerged as a reliable model in medical imaging and
to see if some code works. Using this approach, the dicom diagnostic applications. Traditionally, a radiologist manually
slices of each patient are converted into volume files (.pt) analyzed her MRI or CT images to detect spinal abnormalities.
which further fed as input into the CNN layers Manual interpretation of long dimensional features makes it
difficult to capture or predict exact categories and severities.
3D Image-Volume Files: We know what a 2D image is. In this article, the author introduced Deep His Learning his
This is the most common image captured by a camera. framework to help diagnose his SCI features based on the
However, in some contexts, such as medicine using medical segmentation process. Major contributions are the use of
images, it’s all about dealing with 3D images. This type of filtering algorithms, the development of SSI feature
image is called as ‘volume’. An image consists of a series of extraction, the implementation of highly relevant multi-level
2D slices representing the object under analysis, usually a segmentation methods, and the development of deep learning
human body or an organ. Each layer is represented by an CNN frameworks that integrate segmentation with nonlinear
image, a matrix of pixels (x and y coordinates). The Z SVM and random forest models for prediction. . SCI - Class
coordinate indicates the number. These 3D images are labeling of test data. Efficient deep learning frameworks that
actually a series of images. Each sequence corresponds to a find subsets of significant high-dimensional features in low
slice of the human body. They are called 3D images because time and error rate, but require large amounts of
they provide her three-dimensional view of the human body. computational memory and time, as well as large amounts of
Therefore, load a volume sample of the cervical spine resources to perform operations, is a difficult process.
obtained by CAT scan or CT scan. The combination of all
these DICOM images is the volume. Samir S. Yadav and Shivaji Rao, M. Jadhav [11]
proposed a deep neural network-based medical picture
II. RELATED WORK categorization system for disease detection. This article
J.E. Small et al. [8] proposed a model for Cervical spine explains how to classify pneumonia using a chest X-ray
fracture detection using a convolution neural network. The data set and a convolutional neural network (CNN)-based
proposed system built into 2 stages: One area is the suggestion method. Experiments are used to evaluate the three
level and another is the false positive reduction level. This approaches. These classifiers are linear support vector
network produces a segmented 3D map when trained on machines with transfer learning on two convolutional
segmented scans. Proposed regions are extracted from it and neural network models, local rotation, and exposure-free
proceeded as an input to the second stage of the algorithm. functions. group visual geometry, d. H. VGG16,
This stage of the algorithm categorizes each region as either InceptionV3, and training from scratch of a capsule
positive region or negative region. Two different sets of network. A data pre-processing technique used by all three
features are been extracted from each region and merged for methods is called data augmentation. Data augmentation is
the closing decision. Then it is learned from a multi-layer and a natural technique to improve the performance of all three
classification and categorization head that acquires features of his algorithms, according to development outcomes. In
from the previous or last layer of the 3D segmentation network addition to support vector machines with directed fast
for the suggested region. The model's performance matched rotated binaries (ORBs), robust self-reliant basis functions,
that of a real radiologist on the test set on several calculations and capsule networks, transfer learning is a more useful
and metrics such as accuracy and the F1 score. The CNN classification method for small datasets. Retraining a

Authorized licensed use limited to: Universidad del Valle. Downloaded on October 28,2024 at 18:59:36 UTC from IEEE Xplore. Restrictions apply.
specific function on a new target dataset is essential to training a deep learning model. This page outlines the
enhancing performance in transfer learning. Network fundamental ideas behind DL-Systems, some of the pitfalls to
complexity that is appropriate for the amount of the data set avoid, and instructions on how to spot them.
is the second crucial criterion.
Andrew T. Hale et al. [15] suggested comparing X-ray and
Malinda Vania et al. [12] used convolutional neural CT to identify significant cervical spine injuries in the
networks to automatically segment the spine from CT scans pediatric population. Examination of the cervical spine in
by generating additional class labels, as the number of patients children should be done promptly. Since the spine in
experiencing back problems has significantly increased since childhood is dainty, there is concern about the hazard of
2010. suggested a model for Automated Segmentation Images radiation exposure to the spine amidst CT imaging. Therefore,
produced by this model are crucial for facilitating the experts recommend using her CT scans sparingly. However,
diagnosis of spinal condition and carrying out surgery using CT scans are exact and can produce superior image resolution
computer-assisted surgery network. In this technique, an and can offer the same diagnostic accurateness as CT imaging
algorithm was created to segment the spine from CT scans, in identifying cervical spine injuries in children compared to
and the segmentation results were compared to reference X-rays. Conducted a retrospective study of patients evaluating
segmentations found using other well-known techniques. cervical spine injuries in a level 1 trauma center and concluded
They employed a hybrid strategy that combined complete that CT radiographs are excellent for detecting both clinically
convolutional neural networks and related networks. The significant and non-significant cervical spine injuries in
approach includes training and test data preparation, level-set children. and concluded that it is accurate. They also said they
segmentation based on the histogram, and excess generated were concerned about radiation caused by CT scans.
class labels. The original CT data must go through several Therefore, CT scans are only used in severe cases, especially
segmentation phases in order to get the desired final result, in children.
which is a laborious task. Despite the fact that this model is
defined to deliver good results with a volumetric likeness of Andrew T. Hale et al. [16] suggested comparing X- ray
96%. and CT to identify significant cervical spine injuries in the
pediatric population. Examination of the cervical spine in
Naofumi Tomita et al. [13] designed a deep neural network children should be done promptly. Since the spine in
for automatic CT scan vertebral fracture detection in childhood is dainty, there is concern about the hazard of
osteoporosis. Elderly patients frequently experience radiation exposure to the spine amidst CT imaging. Therefore,
osteoporotic vertebral fractures, which have a substantial experts recommend using her CT scans sparingly. However,
negative impact on their quality of life and socioeconomic CT scans are exact and can produce superior image resolution
status. To avoid more fractures and suffering, early discovery and can offer the same diagnostic accurateness as CT imaging
and diagnosis are crucial. Due to the possibility that they may in identifying cervical spine injuries in children compared to
initially be asymptotic, the majority of these OVF instances X-rays. CT radiographs are very good at detecting both
are incorrectly identified and reported on CT scans. There are clinically significant and non-significant cervical spine
two primary parts to the proposed system. feature extraction injuries in children, according to a retrospective study of
module built on CNN. an RNN module that makes a final patients evaluated for cervical spine injuries in a level 1
diagnostic by adding the extracted features together. To trauma center. and concluded that it is accurate. They also said
achieve the highest believable picture recognition accuracy, they were concerned about radiation caused by CT scans.
deep residual networks (ResNet) are employed for Therefore, CT scans are only used in severe cases, especially
preprocessing and feature extraction from 2D CT slices. in children.
Features were retrieved from each slice and sent into an LSTM
network for RNN-based feature aggregation. The CT scan III. PROPOSED METHODOLOGY
volume is treated as a single label in the classification A solution to the problem of detecting cervical fractures
architecture. In such categorization contexts, the resulting from CT scans has been devised in this research utilizing the
models are prone to learning confounding features, which pytorch framework in Python. Below Fig.3 shows the
might result in inaccurate diagnostic results. proposed system for the data set preprocessing, model
Bradley J. Erickson et al. [14] explained that deep learning building, model training and prediction of results using the
(DL) is a well-liked technique for carrying out a variety of trained model.
important tasks in radiology and medical imaging. Certain DL
models are capable of accurately segmenting organs
(essentially tracing boundaries, allowing volumetric
measurements, or calculating other properties). The eventful
features of image regions can be predicted by various DL
networks. For instance, molecular markers of tissue in the
area, or even auguring markers of tissue in the area, can
indicate whether something is cancerous. Although DL is
simpler to implement than conventional machine learning
techniques, it requires more data and time to process and
analyze the output. Important characteristics are instinctively
recognized, but identifying what those characteristics are
might be challenging. can be disputed. The aforementioned
Fig. 3. Proposed System
study demonstrates in great detail how precise, well-liked, and
significant deep learning models are. Additionally, it claims
that additional data and careful results analysis are needed for

Authorized licensed use limited to: Universidad del Valle. Downloaded on October 28,2024 at 18:59:36 UTC from IEEE Xplore. Restrictions apply.
A. Understanding the data set
The dataset consists of DICOM meta data files and csv
files which are used for labeling or classifying metadata files
as fractured or unfractured. Each patient’s CT meta data
consists dicom images of cervical spine region, each dicom
corresponds to slice of the CT scan. For example, consider a
patient who have CT scan slice of 268. However, the slices
individually cannot be used for the information retrieval. In
order to extract any data for the fracture detection these slices
are combined together into a 3D volume file. The bad scans
are dropped and masks are applied to make all the scans Fig. 6. Number of Fractures
consistent also small deformations are taken care. The below
Fig.4 shows how the slices look. B. Preprocessing the dataset
Before feeding the data into the convolution neural
network model, for each patient in the dataset all the two-
dimensional slices are combined together into a three-
dimensional volume file using the pytorch framework because
it has the main feature of automatic differentiation for the
purpose of creating and training the deep neural networks.
C. CNN Architecture
In this approach the goal is to predict the probability of
fracture for each vertebra in 7 cervical vertebrae which are
denoted by C1, C2, C3....C7 as well as overall probability
Fig. 4. CT scans 2D slices whether there is any fracture in the cervical spine. Therefore,
there will be eight columns in the output also any fractures in
Below graphs of Fig.5 shows the analysis of the training the skull, thoracic spine, ribs and clavicles are ignored.
dataset which depicts the detailed exploration of fractures in CNN Model: The Fig.7 shows the layer structure of the
the vertebrae also the number of fractures in overall patients. deep convolution neural network. The blocks and layers used
It is observed that most of the patients have fracture in C7 to build the model can be clearly observed from the
Ascending order of fractures based on vertebrae: architecture diagram given below.
C7 > C2 > C6 > C5 > C1 > C4 > C3 The developed convolution neural network contains a total
From the Fig.6 it is clear that vertebrae C7 has the of 6 blocks. The first three blocks contain a series of four
maximum number of fractures. Also, several patients have layer: a conv3d layer, ReLu activation layer, maxpool3d layer
more than one fracture. and batch normal 3d layer. The fourth block contain adaptive
average pool 3d layer and a flatten layer. The last but one layer
• Patients with 0 fractures: 1058 is a fully connected layer that consists of linear layer with a
ReLu activation layer.
• Patients with 1 fracture: 623
• Patients with 2 fractures: 239
• Patients with 3 fractures: 64
• Patients with 4 fractures: 26
• Patients with 5 fractures: 7
• Patients with 6 fractures: 2

Fig. 7. CNN Architecture

Finally, the sixth block produces the output with a linear

layer. The model takes in the three-dimensional pytorch
volume files as its input during the training, so 3-D layers are
used to build the network. Table 1 depicts the attributes that
are fed to Conv3D layer. 3D batch normal layers in each of
the three blocks used 16, 32 and 64 number of features
respectively. At the end two linear layers are placed to get the
Fig. 5. Exploratory Data Analysis linear output of size 8.

Authorized licensed use limited to: Universidad del Valle. Downloaded on October 28,2024 at 18:59:36 UTC from IEEE Xplore. Restrictions apply.
 TABLE I. ATTRIBUTES OF CONV3D LAYER IN EACH BLOCK
Layer Attributes
Convolution 3D in_channels out_channels kernel_size
Block 1 1 16 71
Block 2 16 32 3
Block 3 32 64 3

The layers that are used in the convolution neural network

model are conv3d, ReLu, maxPool3D, BatchNorm3D,
AdaptiveAvgPool3D, Fully Connected Layer, at last the linear
layer (output layer). The layers that are used in the convolution
neural network model are conv3d, ReLu, maxPool3D,
BatchNorm3D, AdaptiveAvgPool3D, Fully Connected Layer, Fig. 8. Activation Function
at last the linear layer (output layer) which results in 17 layers
of 6 blocks. Adaptive pooling, specifies the size of the output instead
A convolutional layer is an important component of a of input. The step size and core size are automatically selected
convolutional neural network and is always at least the first to suit your requirements and needs. Batch normalization
layer. Its purpose is to detect the presence of a large number layer’s main function is to accelerate the training of deep
of features in an image received as input. This is done by network by reducing the internal covariate shift. A fully
convolution filtering: The principle is to "pull" a window connected layer governs the relationship between feature
representing the features onto the image and compute the locations and classes in an image. The input table of layer is
convolution product between the features and each part of the the result of the previous layer, so it actually corresponds to
scanned image. The features are then displayed as filters. So a the feature map for a particular feature: high values indicate
convolution layer takes multiple images as input and uses each the position of that feature in the image which is more or less
filter to compute the convolution of each image. The filters accurate depending on pooling. If a feature's location at a
correspond exactly to the features required to be found in the specific area of the picture is distinctive of a particular class,
image. The conv3d in pytorch applies 3D convolution to an the value corresponding to it in the table will be assigned
input signal consisting of multiple input layers where greater weight.
parameters are kernel size, padding, stride, dilation. The Model Training
output of the layer with an input size (N, Cin, D, H, W) is given The converted pytorch volume files are given as input to
in Equation 1. the deep convolution neural network model. The CNN model
contains various layers: conv3d layer is used to extract
, features from the 3D images, maxpool3d layer reduces the
!"# $% size and computation. To the total of 2019 patients, they are
weight , ∗ , divided into two sets and are converted into pytorch files as
they are given as input to the 3D model. A batch size of 4 with
&'(
Equation 1: Equation of conv3D learning rate 0.0001 using 20 epochs the model is trained. The
training ended after 13th epoch as the loss value has exceeded
A pooling layer reduces the number of parameters and the limit. Fig.9 shows the loss value per each epoch.
computations in the network. This makes the network more
efficient and avoids overtraining or overlearning.
The maxpool3d applies three-dimensional max pooling on
an input signal comprised of various input planes. The ReLu
which stands for Rectified Linear Units is a non-linear
function whose equation is provided in Equation 2.
)*+, - - /
max 0, -

Equation 2: ReLU Activation function

It replaces all received input values which are negative as Fig. 9. Model performance during training
zeroes. It acts like an activation function. Fig.8 [17] shows the
function of an activation function. Testing the model
Firstly, the test dataset is used to know the accuracy of the
model. And a new input which is not included in neither train
nor test dataset is provided to model in order to make a
prediction. The trained model follows the same process of
converting slices into volume files and then feeding them to
the deep CNN

Authorized licensed use limited to: Universidad del Valle. Downloaded on October 28,2024 at 18:59:36 UTC from IEEE Xplore. Restrictions apply.
 IV. RESULTS & DISCUSSION
In this work a simple and sophisticated solution has been
developed which can detect whether there is a fracture in
cervical spine region or not. All the existing works are only
providing analysis on the CT scans data but did not gave a
proper solution to it and some provided a solution to check the
fracture in whole cervical region. In this solution is provided
to predict the exact region of the fracture in cervical spine
since it comprises of 7 vertebrae. The developed model
outputs a list with 8 fields which give the probability value (0
to 1) of fracture existence in the C1 to C7 vertebrae and overall
value. When a patient’s CT scans are given as input to the
model, it returns the probability of getting fractured at each
vertebra and the overall fracture probability. The Table 2 Fig. 10. Performance Graph of the Model
shows the structure of results that was returned as output from
the trained model. The model return probability of fracture for Most of the previous works on the cervical spine detection
each vertebrae and a overall probability. were done using the MRI scans which are now seldomly used
technique for medical image diagnosis or either most of the
TABLE II. PROBABILITIES OF FRACTURE previous works have the highest false negative rate due to a
smaller number of fractured data.
C1 C2 C3 C4 C5 C6 C7
Overall V. CONCLUSION AND FUTURE WORK
0.065 0.145 0.032 0.058 0.082 0.161 0.244 0.552
7 9 0 1 5 5 1 2 This work offers a solution for the problem of detecting
Patient id: 1.2.826.0.1.3680043.1001
the fracture in the cervical spine was implemented.
Convolution neural network was used for identification of
It is observed that C7 have the highest probability and is fractures in the CT scans which are stored in the form of
labeled as fractured for the given patients CT scans as input. DICOM metadata. The model is more useful for the cases
Performance Analysis where the patient needs the immediate observation and the
operating procedure and it could take less time than the
The trained model is tested on the validation data which radiologist who approximately take 10 minutes for the
consists of CT data of 14 patients. The model predicts the observation. This model produces the accuracy of 93% and
probability value i.e., value between 0 and 1. A value the output can be generated in less than 2 minutes. In the future
greater that 0.5 is assumed as 1 and value less than 0.5 is work the model is further extended with a web user interface
assumed as 0. The model predicted the true value (nearest where the user can give the CT scans as the input for the web
as mentioned above) for 12 of 14 validation samples. The page which is integrated with the model and can generate the
Table 3 provides the summary of the model prediction on result whether there is any fracture or which vertebrae is
validation data: true positive, true negative, false positive fractured.
and false negative values.
REFERENCES
TABLE III. CONFUSION MATRIX FOR VALIDATION DATA [1] Guez, M., Hildingsson, C., Nilsson, M., & Toolanen, G. (2002). The
Actual label prevalence of neck pain. Acta Orthopaedica Scandinavica, 73(4), 455-
459.
Predicted Fractured Not Fractured
values Fractured 8 (TP) 1(FP) [2] Gerrelts, B. D., Petersen, E. U., Mabry, J. O. H. N., & Petersen, S. R.
(1991). Delayed diagnosis of cervical spine injuries. The Journal of
Not Fractured 0(FN) 5(TN) trauma, 31(12), 1622-1626.
[3] Dowdell, J., Kim, J., Overley, S., & Hecht, A. (2018). Biomechanics
Above graph in Fig.10 is a line graph plot between input and common mechanisms of injury of the cervical spine. Handbook of
clinical neurology, 158, 337-344.
CT scans with patient id is on the x-axis and Probability of
[4] Hattingh, H. L., Michaleff, Z. A., Fawzy, P., Du, L., Willcocks, K.,
fracture is on the y-axis. It can be seen that the output values Tan, K. M., & Keijzers, G. (2022). Ordering of computed tomography
are probability of getting fractured so the probability range lies scans for head and cervical spine: a qualitative study exploring
between 0-1 and the returned probabilities from the model are influences on doctors’ decision-making. BMC health services research,
normalized to nearer values. In this model if the overall 22(1), 1-13.
probability of getting fractured is > 0.5 then it is considered [5] Schaefer, G., Huguet, J., Zhu, S. Y., Plassmann, P., & Ring, F. (2006,
fractures if it is <0.5 it is considered not fractured as most of August). Adopting the DICOM standard for medical infrared images.
In 2006 International Conference of the IEEE Engineering in Medicine
the unfractured scans have probability <0.5. and Biology Society (pp. 236-239). IEEE.
[6] Shen, D., Wu, G., & Suk, H. I. (2017). Deep learning in medical image
analysis. Annual review of biomedical engineering, 19, 221.
[7] Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G.,
... & Chintala, S. (2019). Pytorch: An imperative style, high-
performance deep learning library. Advances in neural information
processing systems, 32.
[8] Small, J. E., Osler, P., Paul, A. B., & Kunst, M. (2021). Ct cervical
spine fracture detection using a convolutional neural network.
American Journal of Neuroradiology, 42(7), 1341-1347.
[9] Salehinejad, H., Ho, E., Lin, H. M., Crivellaro, P., Samorodova, O.,
Arciniegas, M. T., ... & Colak, E. (2021, April). Deep sequential

Authorized licensed use limited to: Universidad del Valle. Downloaded on October 28,2024 at 18:59:36 UTC from IEEE Xplore. Restrictions apply.
 learning for cervical spine fracture detection on computed tomography
imaging. In 2021 IEEE 18th International Symposium on Biomedical
Imaging (ISBI) (pp. 1911-1914). IEEE.
[10] Ahammad, S. H., Rajesh, V., Rahman, M. Z. U., & Lay-Ekuakille, A.
(2020). A hybrid CNN-based segmentation and boosting classifier for
real time sensor spinal cord injury data. IEEE Sensors Journal, 20(17),
10092-10101.
[11] Yadav, S. S., & Jadhav, S. M. (2019). Deep convolutional neural
network based medical image classification for disease diagnosis.
Journal of Big Data, 6(1), 1-18.
[12] Vania, M., Mureja, D., & Lee, D. (2019). Automatic spine
segmentation from CT images using convolutional neural network via
redundant generation of class labels. Journal of Computational Design
and Engineering, 6(2), 224-232.
[13] Tomita, N., Cheung, Y. Y., & Hassanpour, S. (2018). Deep neural
networks for automatic detection of osteoporotic vertebral fractures on
CT scans. Computers in biology and medicine, 98, 8-15.
[14] Bradley J. Erickson, MD, PhD, Panagiotis Korfiatis, PhD, Timothy L.
Kline, PhD, Zeynettin Akkus, PhD, Kenneth Philbrick, PhD,
Alexander D. Weston, BS (2018) “Deep Learning in Radiology: Does
One Size Fit All?” Journal of the American College of Radiology.
[15] Hale, A. T., Alvarado, A., Bey, A. K., Pruthi, S., Mencio, G. A.,
Bonfield, C. M., ... & Naftel, R. P. (2017). X-ray vs. CT in identifying
significant C-spine injuries in the pediatric population. Child's Nervous
System, 33(11), 1977-1983.
[16] Utz, M., Khan, S., O’Connor, D., & Meyers, S. (2014). MDCT and
MRI evaluation of cervical spine trauma. Insights into imaging, 5(1),
67-75.
[17] Stevens, E., Antiga, L., & Viehmann, T. (2020). Deep learning with
PyTorch. Manning Publications.

Authorized licensed use limited to: Universidad del Valle. Downloaded on October 28,2024 at 18:59:36 UTC from IEEE Xplore. Restrictions apply.