168,99 €
Mehr erfahren.
- Herausgeber: John Wiley & Sons
- Kategorie: Wissenschaft und neue Technologien
- Sprache: Englisch
ARTIFICIAL INTELLIGENCE FOR BONE DISORDER The authors have produced an invaluable resource that connects the fields of AI and bone treatment by providing essential insights into the current state and future of AI in bone condition diagnosis and therapy, as well as a methodical examination of machine learning algorithms, deep learning approaches, and their real-world uses. The book explores the use of artificial intelligence (AI) in the diagnosis and treatment of various bone illnesses. The integration of AI approaches in the fields of orthopedics, radiography, tissue engineering, and other areas related to bone are discussed in detail. It covers tissue engineering methods for bone regeneration and investigates the use of AI tools in this area, emphasizing the value of deep learning and how to use AI in tissue engineering efficiently. The book also covers diagnostic and prognostic uses of AI in orthopedics, such as the diagnosis of disorders involving the hip and knee as well as prognoses for therapies. Chapters also look at MRI, trabecular biomechanical strength, and other methods for diagnosing osteoporosis. Other issues the book examines include several uses of AI in pediatric orthopedics, 3D modeling, digital X-ray radiogrammetry, convolutional neural networks for customized care, and digital tomography. With information on the most recent developments and potential future applications, each chapter of the book advances our understanding of how AI might be used to diagnose and treat bone problems. Audience This book will serve as a guide for orthopedic experts, biomedical engineers, faculty members, research scholars, IT specialists, healthcare workers, and hospital administrators.
Sie lesen das E-Book in den Legimi-Apps auf:
Seitenzahl: 479
Veröffentlichungsjahr: 2024
Ähnliche
Table of Contents
Cover
Table of Contents
Series Page
Title Page
Copyright Page
Foreword
Preface
1 Artificial Intelligence and Bone Fracture Detection: An Unexpected Alliance
1.1 Introduction
1.2 Bone Fracture
1.3 Deep Learning and Its Significance in Radiology
1.4 Role of AI in Bone Fracture Detection and Its Application
1.5 Primary Machine Learning-Based Algorithm in Bone Fracture Detection
1.6 Deep Learning-Based Techniques for Fracture Detection
1.7 Conclusion
References
2 Integrating AI With Tissue Engineering: The Next Step in Bone Regeneration
2.1 Introduction
2.2 Anatomy and Biology of Bone
2.3 Bone Regeneration Mechanism
2.4 Understanding AI
2.5 Current AI Integration
2.6 Applying Deep Learning
2.7 Conclusion
References
3 Deep Supervised Learning on Radiological Images to Classify Bone Fractures: A Novel Approach
3.1 Introduction
3.2 Common Bone Disorder
3.3 Deep Supervised Learning’s Importance in Orthopedics and Radiology
3.4 Perspective From the Past
3.5 Essential Deep Learning Methods for Bone Imaging
3.6 Strategies for Effective Annotation
3.7 Application of Deep Learning to the Detection of Fractures
3.8 Conclusion
References
4 Treatment of Osteoporosis and the Use of Digital Health Intervention
4.1 Introduction
4.2 Opportunistic Diagnosis of Osteoporosis
4.3 Predictive Models
4.4 Assessment of Fracture Risk and Osteoporosis Diagnosis by Digital Health
4.5 Clinical Decision Support Tools, Reminders, and Prompts for Spotting Osteoporosis in Digital Health Settings
4.6 The Role of Digital Health in Facilitating Patient Education, Decision, and Conversation
4.7 Conclusion
References
5 Utilizing AI to Improve Orthopedic Care
5.1 Introduction
5.2 What is AI?
5.3 Introduction to Machine Learning: Algorithms and Applications
5.4 Natural Language Processing
5.5 The Internet of Things
5.6 Prospective AI Advantages in Orthopedics
5.7 Diagnostic Application of AI
5.8 Prediction Application With AI
5.9 Conclusion
References
6 Significance of Artificial Intelligence in Spinal Disorder Treatment
6.1 Introduction
6.2 Machine Learning
6.3 Methods Derived From Statistics
6.4 Applications of Machine Learning in Spine Surgery
6.5 Application of AI and ML in Spine Research
6.6 Conclusion
References
7 Osteoporosis Biomarker Identification and Use of Machine Learning in Osteoporosis Treatment
7.1 Introduction
7.2 Biomarkers of Bone Development
7.3 Biomarkers for Bone Resorption
7.4 Regulators of Bone Turnover
7.5 Methods to Identify Osteoporosis
7.6 Conclusion
References
8 The Role of AI in Pediatric Orthopedics
8.1 Introduction
8.2 Strategy Based on Artificial Intelligence
8.3 Several Applications of Artificial Intelligence
8.4 Conclusion
References
9 Use of Artificial Intelligence in Imaging for Bone Cancer
9.1 Introduction
9.2 Applications of Machine Learning to Cancer Diagnosis
9.3 Artificial Intelligence Methods for Diagnosing Bone Cancer
9.4 Methodologies for Constructing Deep Learning Model
9.5 Clinical Image Applications of Deep Learning for Bone Tumors
9.6 Conclusion
References
Index
End User License Agreement
List of Tables
Chapter 1
Table 1.1 Comparison of fractures with diagnosis.
Chapter 8
Table 8.1 Illustration of a patient’s step-wise journey from hospitalization t...
Chapter 9
Table 9.1 Numerous machine learning algorithms’ advantages and limitations.
List of Illustrations
Chapter 1
Figure 1.1 Different types of bone fracture.
Figure 1.2 Deep learning-based work process for fracture detection.
Chapter 2
Figure 2.1 Types of cells that are found within bone tissue.
Figure 2.2 Regenerative bone tissue possibilities. Autograft, allograft, and x...
Figure 2.3 Tissue engineering and biomaterial paradigm for bone repair. After ...
Figure 2.4 Deep neural network concept. The input layer, which has multiple no...
Figure 2.5 AI’s future: clinician-centered. Clinicians continuously train conn...
Chapter 3
Figure 3.1 Diagrammatic representation of intelligent bone fracture detection ...
Figure 3.2 Diagrammatic representation of machine learning in the field of rad...
Figure 3.3 Diagrammatic representation of Haar wavelet image compression.
Figure 3.4 Some examples of photos processed using standard for image feature ...
Chapter 4
Figure 4.1 Solutions using AI that have been developed and are currently avail...
Figure 4.2 Sub-categories of digital health.
Chapter 5
Figure 5.1 The ROC curve is in red, and AUC is the area between red and green....
Chapter 6
Figure 6.1 AI will cause a shift in the spine care industry by enhancing the s...
Figure 6.2 The major subfields of AI, like machine learning techniques, and ho...
Figure 6.3 A simplified support vector machine (SVM) for binary classification...
Figure 6.4 A decision tree that has been trained to predict the likelihood of ...
Chapter 7
Figure 7.1 Biochemical biomarkers of bone turnover. Blue boxes/arrows indicate...
Chapter 8
Figure 8.1 Several radiology departments have begun using artificial intellige...
Figure 8.2 The use of AI to identify wrist fractures in radiographs taken from...
Chapter 9
Figure 9.1 A historical perspective on the evolution of AI, ML, radiomics, DL,...
Figure 9.2 A quick look at the differences between radiomics, traditional mach...
Figure 9.3 (a) An example of MRI. (b) CT and (c) ultrasound images.
Guide
Cover Page
Table of Contents
Series Page
Title Page
Copyright Page
Foreword
Preface
Begin Reading
Index
WILEY END USER LICENSE AGREEMENT
Pages
ii
iii
iv
xi
xii
xiii
xiv
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
Scrivener Publishing100 Cummings Center, Suite 541JBeverly, MA 01915-6106
Publishers at ScrivenerMartin Scrivener ([email protected])Phillip Carmical ([email protected])
Artificial Intelligence for Bone Disorder
Diagnosis and Treatment
Rishabha Malviya
School of Pharmacy, Galgotias University, Greater Noida, India
Shivam Rajput
School of Pharmacy, Galgotias University, Greater Noida, India and
and
Makarand Vaidya
Hindu Rao Hospital, New Delhi, India
This edition first published 2024 by John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, USA and Scrivener Publishing LLC, 100 Cummings Center, Suite 541J, Beverly, MA 01915, USA© 2024 Scrivener Publishing LLCFor more information about Scrivener publications please visit www.scrivenerpublishing.com.
All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any form or by any means, electronic, mechanical, photocopying, recording, or otherwise, except as permitted by law. Advice on how to obtain permission to reuse material from this title is available at http://www.wiley.com/go/permissions.
Wiley Global Headquarters111 River Street, Hoboken, NJ 07030, USA
For details of our global editorial offices, customer services, and more information about Wiley products visit us at www.wiley.com.
Limit of Liability/Disclaimer of WarrantyWhile the publisher and authors have used their best efforts in preparing this work, they make no representations or warranties with respect to the accuracy or completeness of the contents of this work and specifically disclaim all warranties, including without limitation any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created or extended by sales representatives, written sales materials, or promotional statements for this work. The fact that an organization, website, or product is referred to in this work as a citation and/or potential source of further information does not mean that the publisher and authors endorse the information or services the organization, website, or product may provide or recommendations it may make. This work is sold with the understanding that the publisher is not engaged in rendering professional services. The advice and strategies contained herein may not be suitable for your situation. You should consult with a specialist where appropriate. Neither the publisher nor authors shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages. Further, readers should be aware that websites listed in this work may have changed or disappeared between when this work was written and when it is read.
Library of Congress Cataloging-in-Publication Data
ISBN 978-1-394-23088-4
Cover image: Pixabay.ComCover design by Russell Richardson
Foreword
I am delighted to introduce the book “Artificial Intelligence for Bone Disorder: Diagnosis and Treatment,” authored by Dr. Rishabha Malviya. This extensive collection of reports explores the relationship between artificial intelligence and bone problems, illuminating how AI can revolutionize orthopedics and radiology.
It is becoming increasingly apparent as we negotiate the changing landscape of healthcare that new technologies have enormous potential to revolutionize patient care. The integration of artificial intelligence, in particular, into the diagnosis and treatment of bone problems gives an intriguing prospect for improvements in healthcare. Artificial intelligence has emerged as a potent tool in many fields.
This book covers a wide range of issues. It thoroughly investigates how artificial intelligence can improve knowledge and management of bone fractures, tissue engineering for bone regeneration, deep supervised learning for fracture classification, treatment for osteoporosis, orthopedic care, treatment of spinal disorders, pediatric orthopedics, and diagnosis of bone cancer.
This book will be an essential tool for academicians, students, and healthcare professionals. The book offers a road map for utilizing the promise of artificial intelligence to enhance patient outcomes by digging into the complexities of AI algorithms, deep learning techniques, and their application to bone problems.
I appreciate Dr. Rishabha Malviya and all the authors who contributed to this important work with their commitment and knowledge. I am confident that “Artificial Intelligence for Bone Disorder: Diagnosis and Treatment” will inspire additional investigation and progress in the application of AI in bone healthcare, ultimately proving beneficial to patients and medical professionals worldwide.
I am assured that readers will find this book fascinating and informative information that encourages future research, collaborative efforts, and innovations in bone disorder diagnosis and treatment.
Dr. (Prof.) Seifedine Kadry
Professor
Noroff University College, Norway
FIET, SMIEEE, ACM
Preface
Healthcare is one of several sectors that Artificial Intelligence (AI) is revolutionizing. In recent years, AI has demonstrated incredible potential for transforming the diagnosis and management of a wide range of medical diseases. The field of bone problems is one area where AI has a lot of potential.
This book looks into the blending of AI and bone health care. It addresses the various uses of AI in the diagnosis and treatment of bone fractures, tissue engineering, orthopedic care, spinal problems, pediatric orthopedics, bone disease prediction, and imaging for bone cancer through a careful selection of chapters.
Thanks to a thorough analysis of the relationship between bone problems and artificial intelligence, the reader will gain unmatched insight into the most recent developments and prospects. This comprehensive volume proceeds from introductory chapters that provide a foundation in bone fracture detection and tissue engineering to an in-depth examination of the application of AI in particular domains like orthopedics, radiology, and pediatrics orthopedics.
The authors’ combined efforts have produced an invaluable resource that connects the fields of AI and bone treatment. The book provides essential insights into the current state and future of AI in bone condition diagnosis and therapy, as well as a methodical examination of machine learning algorithms, deep learning approaches, and their real-world uses.
May this book spark your interest, inspire your creativity, and motivate readers to take advantage of artificial intelligence’s promise to advance bone health care and ultimately help people worldwide.
The Authors
October 2023
1Artificial Intelligence and Bone Fracture Detection: An Unexpected Alliance
Abstract
Bone fractures are the most frequent type of injury sustained in accidents. Clinicians mostly use radiographs and CT (computed tomography) scans to diagnose fractures, yet it is frequently impossible to make a correct diagnosis only from images. Also, a lack of doctors in areas where healthcare is scarce, a lack of specialized medical staff in overpopulated hospitals, or pressure caused by a high workload may all make it more likely that a fracture will be misdiagnosed or not heal well. “Artificial intelligence” (AI) is the process of programming computers to act like smart people with little or no human interaction. Using computer vision and AI techniques like deep learning and machine learning for image processing is becoming more and more important for recognizing bone fractures. This chapter begins by discussing bane fracture and its different kinds. Consequently, the function and uses of AI in bone fracture are described. This chapter also discusses various algorithms based on machine learning and deep learning and their importance in bone fracture detection.
Keywords: Artificial intelligence, bone fracture, deep learning, machine learning, radiology, algorithm, fracture detection
1.1 Introduction
The term “artificial intelligence” (AI) is used to describe the practice of utilizing computers to simulate intelligent behavior with little or no input from humans [1]. In addition, recent advances in AI, and notably machine learning and deep learning, have allowed machines to significantly improve their ability to encode and interpret imaging data [2]. In medical science, AI is used for a wide variety of tasks, including symptom identification and radiological im age processing. Machine learning is the iterative process that artificial intelligence utilizes to improve in response to the training technique, further enhancing the outputs for judgments and predictions. The learning process involves various types of learning [3]. One type of machine learning, supervised learning, trained itself to anticipate future outcomes using a set of known data [4]. To categorize newly entered data, a trained model is then applied to the data using a classification technique [5]. The resulting groups are then used to make judgments. Unsupervised learning methods are another option that can be utilized to generate judgments in situations where trained models are not available [6]. Data analysis, computation, and application are all part of the learning process [7, 8]. Training in data collection is a common definition of learning. Like how training is a combination of numerous techniques, deep learning is based on supervised and unsupervised learning as well as the use of networks like artificial neural networks. Artificial intelligence is a branch of computer science that automates the process of combing through enormous datasets in search of recurring patterns and trends, frequently in datasets that are utterly unrelated to one another, like medical imaging for the detection of fractures. The application of machine learning and deep learning techniques has significant advantages in the disciplines of image processing, computer visualization, and medicine.
Layered artificial neural networks are the building blocks of deep learning, a subfield of machine learning. Each layer is composed of a predetermined number of units, each of which is a condensed representation of a neuron cell based on the anatomy of neurons seen in the human brain [9]. In specialized tasks, deep learning algorithms can match and occasionally even surpass human performance [10, 11]. Information technology has been revolutionized by deep learning, which has made it possible to quickly solve large-scale, data-driven challenges. The automated classification and detection of proximal humerus fractures using a deep learning algorithm is a prime example of artificial intelligence (also known as machine intelligence) in action, as it uses numerous layers of network systems to obtain features to enhance the quality of perception and accuracy. Deep learning is essential for the computation and processing needed by artificial intelligence. Deep learning uses neural networks, which are made up of input layers and hidden layers, to carry out the necessary computations and processing [12]. As a result, deep neural networks have emerged as a promising tool for improving the efficiency and precision of medical picture diagnosis. Deep learning has been gaining popularity over the past few years [13]. Recent studies have shown that deep learning can do difficult interpretation at the same level as human medical professionals [14, 15]. In the field of orthopedic traumatology, several studies have been carried out using deep learning in radiography to diagnose fractures [16]. However, research on the application of deep learning to CT imaging of fractures is few.
Any kind of bone break, from the tiniest hairline to the most severe severing, is considered a fracture. Fractures can be either closed or open, depending on the severity of the damage. The amount of force required to cause a fracture varies [17]. Both the patient’s age and the affected bone region play a role in the degree of severity. Orthopedics is the medical specialty concerned with damaged bones, also known as fractures. Bone fractures can occur in any orientation after being broken by an outside force or stimulation, including but not limited to the vertical, horizontal, and oblique planes. Failure to diagnose a fracture in a clinical setting where specialist knowledge is lacking can have serious effects and reduce the likelihood of a positive result.
In most cases, an X-ray can detect and identify any form of fracture. Fracture confirmation via computed tomography (CT) or magnetic resonance imaging (MRI) can be performed later. X-rays are an electromagnetic radiation-based imaging technology used to provide images of structures beneath the skin. The patient is positioned properly for X-ray imaging of the affected area. Computer tomography, more commonly known as a “CT scan,” is a medical imaging technique that employs a combination of a computer as well as an X-ray machine to produce cross-sectional images as well as three-dimensional reconstruction visuals to improve diagnostic capability and image clarity. To create images, magnetic resonance imaging, often known as MRI, makes use of the paired as well as unpaired magnetic fields of the proton nucleus of hydrogen atoms in water, which is prevalent throughout the body. Even if a fracture has not yet caused any noticeable symptoms, an MRI can help find it—for example, stress-related fractures as well as fractures that occur in situations where imaging procedures such as X-rays and CT scans are dangerous to perform during pregnancy. As a result of recent developments in machine learning, it is now possible to employ computers to assist with medical diagnosis, ushering in a veritable renaissance in the scientific community. It uses up-to-date, cutting-edge methods and tools to give a rapid diagnosis with increased precision. If a fracture is not detected on an X-ray, the patient could suffer serious consequences that hinder their rehabilitation.
1.2 Bone Fracture
There are various ways to categorize fractures. Examples of traumatic fractures include those sustained in motor vehicle accidents, physical altercations, train mishaps, slips and falls, and similar occurrences. Bones that have been compromised by cancer metastasis (bony metastasis) or osteoporosis (increased porosity of the bone) are more likely to suffer a pathological fracture. Implant-related fractures occur at mechanical weak points brought on by persistent pressure. Children and teenagers may only have a partial fracture line since their bones are still growing. These fractures, often known as “greenstick fractures,” are incomplete due to the failure of one tension cortex. Children may experience a deformation of the bone without a clear fracture line. Numerous other factors, such as fracture pattern, fragment displacement, dislocation or subluxation of adjacent joints, quantity of pieces, etc., can be used in the classification process. The following Table 1.1 summarizes the common breakdowns and compares them across several dimensions.
Figure 1.1 depicts the various types of fractures that might occur. The first type of fracture is known as a stable and closed fracture, and it occurs when the bone fragments are ruled into a straight line and are just slightly out of place. The second type of fracture is known as an open or compound fracture, and it occurs when a bone breaks and protrudes completely through the skin. The bone may be visible in certain circumstances and hidden in others. Afterward, a transverse fracture occurs. An example of a transverse fracture is one in which the fracture line runs at an angle to the body’s longitudinal axis. Next is oblique fractures, also called multiple fractures since they occur at different angles. Furthermore, a comminuted fracture occurs when the bone breaks into more than two pieces. Subsequently, a condition known as greenstick fracture occurs when bone fractures but do not fully separate. Kids may feel this way. Finally, spiral fractures occur when a break in the bone progresses in a spiral pattern.
However, only X-rays can correctly identify a fracture [18], even though a visual deformity, particularly a large one, may be discernible to the human eye. To provide a comprehensive description of a fracture, the following information must be provided: the name of the bone, the region of the bone, the pattern of the fracture line, the presence or absence of compression, the existence or absence of displacement of fracture fragments, the type and degree of displacement, the presence or absence of any pre-existing pathology as well as the existence or absence of any associated joint pathology, such as dislocation or subluxation. This information can be found in Table 1.1.
1.3 Deep Learning and Its Significance in Radiology
The term “deep learning” (DL) is utilized to designate a collection of approaches employed in the discipline of machine learning and the much broader subject of artificial intelligence. Both fields are subfields of the field of artificial intelligence. As a subfield of AI, deep learning is characterized by its reliance on simulated neural networks. The units, which are based on the architecture of the human brain, increase in number with each successive layer. The idea of deep learning revolutionized the field of information technology (IT) by providing large-scale answers to problems that had previously taken hours to resolve. The common goal of these algorithms is to comprehend the information at hand. For semantic image processing applications, deep learning has proven to be particularly effective [19].
Table 1.1 Comparison of fractures with diagnosis.
Type of fractures
Symptoms
Diagnosis
Consequences if missed
Possibility of correction
Stable and closed
Pain, swelling, loss of transmitted movements, slight or no deformity
X-rays
Angulation, pain
Yes
Open and compound
Pain, bleeding, open wound, loss of transmitted movements, crepitus, deformity
X-rays
Malunion, infection, non-union, septic arthritis, osteomyelitis, Volkmann’s ischemia, compartment syndrome
Less
Transverse
Pain, deformity, swelling, crepitus, loss of transmitted movements
X-rays
Angulation, persistent, stiffness, deformity, malunion, shortening, non-union
Less
Oblique
Pain, deformity, swelling, crepitus, loss of transmitted movements
X-rays
Angulation, stiffness, persistent deformity, malunion, non-union, shortening
Less
Comminuted
Gross swelling, total loss of transmitted movements, severe pain, crepitus, gross deformity
X-rays
Compartment syndrome, neurovascular injury, Volkmann’s ischemia, subdeck’s osteodystrophy malunion, nonunion
Maybe or may not
Greenstick
Pain, deformity, swelling, restricted mobility, angulation
X-rays
Angulation, stiffness persistent deformity, malunion
Yes
Spiral
Pain, deformity, swelling, loss of transmitted movements
X-rays
Deformity, angulation
Yes
Figure 1.1 Different types of bone fracture.
Classification, segmentation, and noise reduction in medical images are just some of the many applications of DL approaches in use today. What is more, DL algorithms can learn from previous successes by using a technique known as transfer learning [20]. The significance of deep learning in radiology cannot be overstated. As a result, doctors will be able to give their patients better care and more accurate diagnosis. In the field of fracture diagnosis and localization on medical images such as radiographs and CT scans, deep learning achieves better results than humans under the current conditions [21].
1.4 Role of AI in Bone Fracture Detection and Its Application
As early as 1960, AI was being used to analyze X-ray pictures for medicinal purposes [22]. The data was initially converted to a numerical format for computation and analysis. Computer-aided detection was a later development in AI. Improvements in computational speed can be traced to the maturation of a system’s ability to learn from past mistakes; this ability is known as machine learning. The advancement of AI systems in many fields is made possible through machine learning. AI models that undergo training and learning yield outcomes that are both more accurate and more efficient when applied to datasets [23, 24]. Artificial intelligence has proven very effective at sorting urgent situations from less urgent ones.
Bone fractures in X-rays can be identified by artificial intelligence systems that have been trained using machine learning methods. AI-enabled models can provide automatic diagnoses of bone fractures. When applied to X-ray pictures, a neural network that has been trained with appropriate data yields reliable findings. In addition, AI and, more specifically, deep learning, which includes neural networks, enhance the medical research area by allowing for enhanced visualization of imaging data. Bone fracture detection architecture development relies heavily on convolutional neural networks. It can produce results quickly and accurately.
Several studies have demonstrated that deep learning may be utilized to diagnose fractures [25] by undertaking a retrospective analysis to investigate the relevance of transfer learning using deep convolutional neural networks (CNNs), which had been pre-trained on photos that were not related to medicine, to the job of automating fracture recognition on plain wrist radiographs. The authors utilized the inception V3 convolutional neural network [26], which was trained for the IMageNet Large Visual Recognition Challenge on images other than radiography [27]. To complete the binary classification task, the top layer of the Inception V3 network was retrained with the help of a training dataset consisting of 1,389 radiographs annotated by humans. They were successful on the test dataset, with an area under the curve (AUC) of 0.95 (139 radiographs). This proved that the problem of fracture diagnosis on plain radiographs may be effectively applied to a CNN model pre-trained on non-medical pictures. Both the specificity and sensitivity were maximized at 0.90 and 0.88, respectively. Automated fracture analysis techniques like segmentation, edge recognition, and feature extraction have never been as accurate as they are now (prior studies only reported sensitivities and specificities in the 80%–85% range). There are still questions that need to be answered, even though this study proves the premise. At the end of the training phase, it was revealed that the accuracy of the training data and the accuracy of the validation data had a very minor but noticeable difference. Most likely, the problem was caused by excessive tailoring. There are a variety of approaches that can be taken to cut down on overfitting. It is possible to make the process of training more effective by eliminating the influence of attributes that are not important with automated segmentation of the most pertinent region of interest. One such strategy that could assist in lowering the incidence of overfitting is the application of cutting-edge augmentative approaches. In the realm of machine learning, another barrier that frequently arises is the size of the research sample [too small (1,000:10,000)]. The higher the size of the population that is being sampled, the more accurately it will reflect the overall population [28].
The researchers also employed plain AP (perfect anteroposterior) shoulder radiographs to investigate the effectiveness of deep learning in detecting and categorizing proximal humerus fractures. This was done by comparing the radiographs to a database of known fractures. The conclusions of the CNN network were contrasted with those of several specialists (general physicians, radiologists, and orthopedic surgeons). They used a pre-trained ResNet-152 model that was modified specifically for their proximal humerus fracture datasets, which included a total of 1,891 plain AP radiographs. The trained CNN had outstanding ability in distinguishing between normal and broken proximal humeri. On straightforward AP radiographs of the shoulder, the results for the classifying fracture type were also positive. The CNN performed better than primary care physicians and general orthopedic surgeons, and it was on par with shoulder specialists in terms of its effectiveness. This suggests that other fractures and orthopedic illnesses that can be accurately diagnosed with plain radiographs, such as those of the proximal humerus, could benefit from automated diagnosis and classification in the future [36].
The researchers carried out a study to test deep learning’s capacity to automatically identify osteoporotic vertebral fractures (OVFs) on CT scans [29]. This evaluation was carried out as part of a retrospective study. The authors developed a machine learning system that relied solely on a deep neural network framework for all its functionality. Their OVF detection method consisted of two primary components: (1) a CNN-based feature extraction module and (2) an RNN (recurrent neural network) module that aggregated the retrieved features and produced the final diagnostic. Both modules were trained using the same data. A deep residual network, abbreviated as ResNet, was utilized to analyze the CT scans and extract the features [30]. For their training dataset, the researchers made use of 1,168 CT scans; for their validation set, they made use of 135 CT scans; and for their test dataset, they made use of 129 CT scans. On an independent test set, their system achieved levels of accuracy and F1 scores (mean of precision and recall) that were equivalent to those of expert radiologists. By eliminating the need for human intervention during OVF screening, this computerized detection system may help reduce false-negative errors in the diagnosis of asymptomatic, early-stage vertebral fractures.
1.4.1 Data Pre-Processing
The dataset can be pre-trained with non-medical pictures and then used for the task of automating fracture identification in plain radiographs. Preprocessing a dataset of broken photos can enhance data attributes and hide undesired aspects [31]. Pre-processing the data for a learning model takes a considerable amount of time. The goal here is to get the data ready for analysis and calculations. Overfitting is a potential issue during data retrieval and selection [32]. Overfitting occurs when the data absorbs irrelevant information, such as noise, during the learning process. The learning process itself takes considerable time, and the initial dataset acquisition is just the beginning. The next step is a labeling procedure that must be carefully supervised. This information is then separated into two groups: fractured and unbroken. The fracture’s identification was verified by a surgeon with more than 3 years of experience. An orthopedic specialist’s expertise is used to interpret a dataset of X-ray images obtained from a specialized hospital.
1.5 Primary Machine Learning-Based Algorithm in Bone Fracture Detection
Early fracture detection mechanisms relied on the neck-shaft angle (NSA) alone to function. As an example, if the NSA value is less than 116, a fracture is assumed to be present in the image. The mistake rate is close to 8% since the single feature is insufficient for identifying fractures. Therefore, researchers present a novel classification-based approach to detecting mild diseases by analyzing dragging characteristics in femur X-rays and performing the textural assessment of the bony trabecular pattern.
The process of feature extraction is crucial for using foundational machine learning methods for fracture identification. Gabor filters, Markov random field, and intensity gradient direction are techniques used for feature extraction from femur X-rays [33, 34]. Some researchers have exclusively used the Gabor filter for feature extraction while using classifiers for fracture identification [35]. Some scientists employ a gradient transformation using homotopic functions. Regression trees and SVM have several potential applications, including medical diagnosis and weather forecasting [36, 37]. It also expresses one take on the central theme and the supporting issues, disputes, challenges, etc., by analyzing the article’s topic considering the overarching theme of the publication and comparing the results to previous or ongoing research or practice in the field.
1.5.1 Ensemble-Based Classification System
To enhance predictions, ensemble approaches combine multiple algorithms into a unified framework. Associating different kinds of learning models helps to refine the results. Since this is the case, ensemble methods are favored over alternative approaches. Ensemble-based classification systems are collections of classifiers that work together to provide a composite output [38]. Unlike the other techniques, which utilize the same learning method for all the datasets, the initial approach to the ensemble methodology is based on applying multiple learning methods on separate training sets. For accuracy, the classifier system combines the results of a small number of base classifiers [39].
To pick the best possible subset of classifiers, ensemble classifiers make use of selection procedures. There are two distinct ways that one can decide between two options: The most efficient classifier would be chosen to be used on the validation data in an entirely static setup. Both dynamic classifier selection and dynamic ensemble selection are valid applications of the dynamic methodology. Dynamic classifier selection uses a set of preselected classifiers as a foundation upon which to build an ensemble. The goal of dynamic ensemble selection is to find the most predictive ensemble combination [40]. Bagging and boosting are two examples of ensemble methods that have medical applications for fracture identification [41].
To better diagnose osteoporotic fractures, researchers have turned to ensemble-based learning techniques like bagging, boosting, and random subspaces [42]. Classifiers are developed from many models, each with its own unique set of properties. Bone fracture detection uses the ensemble method of 10-fold cross-validation [43]. In the research process, feature selection often comes first, followed by using a machine learning strategy. This method has the benefit of preventing additional, redundant bone testing.
1.5.2 Bagging
Bagging is a popular method that uses a single dataset to train many models. Some typical and recurring occurrences are the result of randomly selecting the initial functionality of a distinct dataset from the same dataset k time. Training several related datasets in parallel has many benefits, including reducing training time and enhancing the quality of results for tasks like decision trees that use many datasets. Joining the results of multiple classifiers where the majority value is the same increases precision. It relies solely on the arithmetic mean of results from the various basic classifiers. The bagging technique calculation formula is shown in Equation 1.1 as follows:
Algorithm
Input:
D: dataset
C: number of classifiers
S: classification Scheme
Output:
Predicted output by Ensemble—a classifier model
Method:
1. For i to C do // create C models
2. Construct dataset Di from D
3. Apply Di with Ci and generate model
4. End;
5. Conjoint the outcomes in Ensemble S
6. Generate predicted value
Instead of randomly selecting from the training set, the authors advise using a method known as “wagging (weights aggregation),” which can be found in [44]. This tactic constantly rearranges the data collection, which makes it much simpler to identify any recurring trends or patterns. During the process of weighing, extra characteristics were assigned to each weight in the calculation. These characteristics included a random Gaussian distribution with a set standard deviation and a mean of zero for each weight. First, for testing, noise was introduced to the weights of the examples during each iteration of the technique. Next, a single classifier was induced when the operation was completed.
Bagging classification is the method that is utilized for fracture detection [45]. After labeling the X-rays and using filters to cut down on noise, the authors take a series of images using the X-ray machine. In addition to that, the bagging method is utilized as a classification strategy so that a basic classifier can be developed.
1.5.3 Boosting
Boosting is an alternative method that employs the same learning mechanism over numerous datasets to achieve its desired results. It is an example of an iterative approach. The accuracy of the classifier can be enhanced by applying weights to each instance, which is accomplished by the dynamic partitioning of training sets. After the fundamental classifiers have been developed and put to the ensemble, the first step is to reweigh each instance individually. The formation of a single prediction requires the combination of several weights derived from various base classifiers. The final prediction is the weighted average of all the predictions made by the individual base classifiers or the “conjoint.”
The results from many base classifiers are averaged and weighted as shown in Equation 1.2 to produce the best results for boosting.
The Adaptive Boosting (AdaBoost) algorithm is an example of a boosting ensemble technique. In this research, the researchers introduced MultiBoosting, an algorithmic approach to decision-making that is an extension of the AdaBoost method [46]. Greater precision in medical diagnosis may be achieved with the use of gradient boosting machines (GBM) [47]. It has been discovered that GBM produce reliable fracture predictions. Before feeding data into a GBM model, regression methods are used. For fracture prediction, a unique model is built with the R package.
1.5.4 Stacking
When multiple learning classifiers’ results are combined into a single classifier, the resulting prediction exhibits less bias and variance error. These methods fall under the umbrella of ensemble learning. Stacking involves dividing a problem space into smaller pieces that may be handled by various kinds of learning algorithms. For each distinct problem partition, a unique learning model is constructed. The combined result is then used as a prediction.
1.5.5 Random Forest
At its core, it relies on the idea of categorizing attributes into different groups. It accomplishes this goal by applying separate feature subspaces to the same dataset used for training. The random forest can produce ensemble classifiers since it uses a huge number of trees. Classification using this method entails constructing several decision trees. The random forest method as well as the feature extraction procedure make it a breeze to perform an analysis of X-ray images [48]. The findings are based on tests that were carried out on a total of 145 X-ray scans. The method that has been suggested can be executed with the assistance of the feature extraction process, the random forest method, as well as the support vector machine method.
1.6 Deep Learning-Based Techniques for Fracture Detection
Fracture identification and characterization are two areas where deep learning could be useful in radiology, as evidenced by the instances (Figure 1.2). Overall, adding deep learning to current radiology procedures has the potential to increase the efficiency and precision of diagnostic tests while simultaneously reducing labor costs by relieving radiologists of some of their more tedious responsibilities. In addition, the algorithms used in deep learning are susceptible to problems such as inter- and intra-variance, both of which can be problematic for human-based diagnosis. Deep learning, when applied in the context of academic research, can identify and classify fractures on plain radiographs and CT scans at least as well as people can.
When using deep neural networks, accurate results can be achieved by merging the predictions generated by many training models [49]. They allow for greater data flexibility and scalability, and they offer new weight sets with each iteration of deep learning, leading to unique predictions. The researchers also talk about a method called Adaptive Differential Parameter Optimization that uses artificial neural networks to identify bone fractures by the separation of fractured and unfractured lines [50]. The approach uses a probabilistic Hough transform for the data preprocessing and a neural network for the classification in order to differentiate between X-rays that contain fractures and those that do not have fractures. The area under the curve was 0.8149, while the overall rate of success was 74.4%.
Figure 1.2 Deep learning-based work process for fracture detection.
A study was conducted to detect the wrist fracture using a convolutional neural network [28]. There are a total of 1,35,845 X-rays, each of which represents an image of a different portion of the body; however, only 34,990 were selected as they were of the wrist bone. From a total of 34,990 X-rays of the wrist, two test datasets were generated, with the first including 3,500 radiographs. There were 1,400 X-rays in the second dataset. There are two stages to the model’s training process. It is first pre-trained, then honed and polished, to detect breaks in the wrist. The researchers generate an AUC curve with a sensitivity of 93.9% and a specificity of 94.5% and then compare these numbers across datasets to determine our performance. Specifically, the researchers use DenseNet, a specific type of neural network architecture, to analyze over 53,000 pelvic X-ray images [51]. Following the application of labeling, a 172-layer deep neural network is used, with each layer featuring 12 features/units. The model’s receiver operating characteristic (ROC) curve boasts an accuracy range from 95% to over 99%.
Transfer learning, a specific type of deep convolutional neural network, is used to detect breaks in X-ray pictures [52]. As a first step, the tool was trained on a dataset of 11,112 photos, from which it correctly labeled 695 fracture images and 694 non-fracture images. By determining a sensitivity of 0.9 and a specificity of 0.88, an AUC value of approximately 0.954 may be generated, which is used in the performance evaluation process. An automated method for detecting and classifying proximal humerus fractures using a deep learning algorithm [53] has been developed utilizing artificial neural networks. The 1,891-image training dataset is a convolutional neural network. Images undergo manual square cropping and resizing to 256 × 256 pixels as part of the data pre-processing phase. Following this, a deep CNN model is implemented, and the data is split into 10 subsets, one of which is utilized as the actual dataset itself, while the remaining nine are used for model training. The model has a predictive accuracy of 0.996, sensitivity of 0.99, and specificity of 0.97. The area under the curve (AUC) is also 0.996.
Bone fracture diagnosis is greatly aided using deep learning [54]. Using a neural network of 1,024 neurons, X-rays of different bodily areas are classified as either fractured or unfractured in order to achieve this goal. The methodology was tested on 30 sample datasets and then applied to 70 datasets. Two of the most common types of pre-processing methods are the Haar wavelet transform and symmetric interferometric feature extraction. The procedure known as scale-invariant feature transform is a multi-stage process that comprises the extraction and orientation of features. This strategy makes use of deep learning in the form of a neural network, and it possesses accuracy of approximately 94%.
The researchers can now identify hip fractures in X-rays by employing a deep neural network [55]. The first is a dataset that has been pre-trained, and it consists of 25,505 limb X-rays (elbows, ankles, wrists, and feet). These datasets are used to train a deep convolutional neural network. In addition, 3,605 X-ray pictures make up the frontal pelvic dataset. In addition, the researchers have a dataset consisting of 100 pelvic X-rays taken for a separate evaluation. To begin, the dataset is preprocessed by having all photos resized to 512 × 512. Fracture identification is modeled using Densenet [56]. Both accuracy and AUC might vary widely depending on the data used. When first used, the pre-training stage achieves an accuracy of 99.5%. A direct application to hip fracture imaging, however, improved the accuracy to 90% and area under the curve to 0.98. In addition, the accuracy was determined after a questionnaire was given to four orthopedic surgeons and two radiologists, with a mean accuracy of 87.7% sensitivity and 98% specificity.
Deep learning algorithms can be broken down into two groups: single- and two-stage classifiers. Two-stage classifiers produce regions that could contain items. Then, a neural network assigns categories to these areas. One-stage detectors do both object localization and classification simultaneously, skipping the region proposal phase. Therefore, single-stage classifiers are quicker than multi-stage ones. When it comes to deep learning-based fracture detection algorithms, the Region-Based Convolutional Neural Network (R-CNN) and the You Only Look Once (YOLO) series are two of the more well-known options. The R-CNN series is less efficient in detecting changes than the YOLO series. Its performance in detecting fractures in real time falls short under realistic conditions.
1.6.1 R-CNN Series
The R-CNN is the pioneering deep learning-based object detection approach. One can use variations of R-CNN that are faster, such as Faster RCNN or Faster R-CNN. Since R-CNN requires around 47 s to analyze each test image, it is not feasible for use in real time. When using slow R-CNN, the regions are only partially sent to CNN, but when using rapid R-CNN, the complete image is fed to the CNN so that it can generate a feature map. It is then common practice to square off the feature maps. Finally, the fully connected layer receives the converted regions of interest (ROls) of a specified size from the ROl pooling layer [57]. If one wants faster results, then that individual can forget about selective search; it is a time-consuming, laborious process that will only slow down that person.
A convolutional neural network is used to analyze the input image and build a feature map when doing faster R-CNN, which is very similar to fast R-CNN. The different network is used to estimate region proposals rather than utilizing a selective search strategy on the feature map to find them. After a ROl pooling layer classifies the picture within the proposed zone and estimates the offset values for the bounding boxes, the proposed regions are shrunk [58].
1.6.2 YOLO Deep Learning Algorithm
Using a convolutional neural network, YOLO can detect objects with a single step. Its speed and efficiency have made it famous. There are several deep learning algorithms out there, but none of them can perform object detection in a single pass. There are a lot of deep learning methods out there, but none of them can do 100% accurate identification in just one go. Due to its ability to detect objects with just a single forward propagation via a neural network, YOLO outperforms any multi-step approach and is thus well suited to usage in real-time contexts. YOLO’s central idea is to feed the network the entire image as input and then have it provide the location of the bounding box and the category to which it belongs as output [59].
The predictions and confidences for each of YOLO’s five bounding boxes are based on the grid unit at the box’s center and are derived from the image’s features. In the YOLO basic frame, w and h were the projected width and height of the full image. You Only Look Once consists mostly of three parts as follows:
Backbone: a convolutional neural network that combines and generates a wide variety of visual features at multiple levels.
Neck: a collection of network layers that, before transferring the information to the prediction layer, combine and mix the attributes of the images being processed.
Head: the head can classify information, recognize bounding boxes, and anticipate the appearance of certain image components. The level of confidence reflects the precision with which the classification was made in light of the circumstances.
Within the YOLOv2 system, a newly developed training algorithm has been incorporated [60]. The k-means clustering method is utilized so that the bounding boxes can be grouped for the clustering process. As a result of the fact that enhancing the intersection over union is the primary objective of the process of developing the a priori box, the IOU value is used in the cluster analysis in the capacity of a distance indicator. It has a very significant improvement in accuracy and recall rate in comparison to YOLOv1. YOLOv3 makes use of both the more advanced basic classification network known as ResNet as well as the classifier known as Darknet53 [61]. A network topology that is analogous to an FPN is utilized in the process of simultaneously performing multiscale prediction. The percentage of false positives for backdrop detection has been significantly reduced, while there has been an increase in both accuracy and speed. YOLOv4 replaces the backbone network with CSPDarknet53 and employs spatial pyramid pooling (SPP) to enhance the receptive area, with PANet serving as the neck. This is done while maintaining the head that YOLOv3 provided. The structure of CSPNet makes it possible to obtain additional information on gradient combinations while simultaneously increasing processing performance. The many feature layers are wholly incorporated into the PANet topology, which considerably improves the faulty feature extraction [62].
1.6.3 YOLOv5
In the year 2020, Glenn and his team introduced the most recent version of the YOLO algorithms, which was called YOLOv5. The Darknet framework was utilized in the production of YOLO models. An organization by the name of Ultralystic adapts older versions of YOLO to the popular deep learning structure known as PyTorch, which is founded on the programming language Python. YOLOv5 had significant engineering improvements over the earlier versions of the software. Python, not C, is being used for the development of YOLOv5. Because YOLOv4 and YOLOv5 are written in two different languages and run on two different frameworks, it is challenging to evaluate how well the two versions perform in comparison to one another. Despite this, YOLOv5 has proven to be more effective than YOLOv4 in several scenarios and has earned a measure of credibility in the area of computer search [63].
The architecture of YOLOv5: YOLOv5 was different from the versions that came before it. PyTorch is utilized here in its place of Darknet. It utilizes CSPDarknet53 as its primary backbone. This backbone manages the recurring gradient information in big backbones, and the feature map contains gradient change. As a result, inference is carried out more quickly, accuracy is improved, and model size is reduced as a result of lower parameter values. It does this by utilizing a framework known as PANet, which speeds up the flow of information. PANet employs a brand new FPN that has a lot of levels. The model’s transmission of lower-level features is improved as a result of this [64].
1.7 Conclusion
Overall, it is possible to conclude that bones play a crucial role in the structure of the human body. In the modern world, artificial intelligence is a cutting-edge idea in the study of medicine. There could be a surge in medical science as a result. The current chapter presents a thorough analysis of the use of artificial intelligence to detect bone fractures. Many reasons, such as an unhealthy lifestyle, the prevalence of traffic accidents, and other factors, have contributed to an increase in bone fractures during the past several years. Artificial intelligence must advance to be used in therapeutic settings such as bone fractures. More effort will lead to better performance, which will aid medical personnel in identifying the fracture.
References
1. Hamet, P. and Tremblay, J., Artificial intelligence in medicine.
Metabolism
, 69, S36–40, 2017 Apr 1.
2. Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L.H., Aerts, H.J., Artificial intelligence in radiology.
Nat. Rev. Cancer
, 18, 8, 500–10, 2018 Aug.
3. Han, J., Pei, J., Tong, H.,
Data mining: Concepts and techniques
, Morgan Kaufmann, Burlington, Massachusetts, United States, 2022 Jul 2.
4. Ilango, V., Subramanian, R., Vasudevan, V., A five step procedure for outlier analysis in data mining.
Eur. J. Sci. Res.
, 75, 3, 327–39, 2012.
5. Veni, C.K. and Rani, T.S., Ensemble based classification using small training sets: A novel approach, in:
2014 IEEE Symposium on Computational Intelligence in Ensemble Learning (CIEL)
, IEEE, pp. 1–8, 2014 Dec 9.
6. Mishra, D. and Soni, D., A comprehensive overview on data mining: Approaches and applications.
Int. J. Comput. Sci. Inf. Technol.
, 5, 6, 7814–6, 2014.
7. Koteeswaran, S., Visu, P., Janet, J., A review on clustering and outlier analysis techniques in datamining.
Am. J. Appl. Sci.
, 9, 2, 254, 2012.
8. Li, J., Roy, P., Khan, S.U., Wang, L., Bai, Y., Data mining using clouds: An experimental implementation of apriori over mapreduce, in:
12th International Conference on Scalable Computing and Communications (ScalCom)
, pp. 1–8, 2012 Dec.
9. McCulloch, W.S. and Pitts, W., A logical calculus of the ideas immanent in nervous activity.
Bull. Math. Biophys.
, 5, 115–33, 1943 Dec.
10. Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G., Petersen, S., Human-level control through deep reinforcement learning.
Nature
, 518, 7540, 529–33, 2015 Feb.
11. Moravčík, M., Schmid, M., Burch, N., Lisý, V., Morrill, D., Bard, N., Davis, T., Waugh, K., Johanson, M., Bowling, M., Deepstack: Expert-level artificial intelligence in heads-up no-limit poker.
Science
, 356, 6337, 508–513, 2017 May 5.
12. Johari, N. and Singh, N., Bone fracture detection using edge detection technique, in:
Soft Computing: Theories and Applications: Proceedings of SoCTA 2016
, vol. 2, pp. 11–19, Springer Singapore, 2018.
13. Adams, M., Chen, W., Holcdorf, D., McCusker, M.W., Howe, P.D., Gaillard, F., Computer vs human: Deep learning versus perceptual training for the detection of neck of femur fractures.
J. Med. Imaging Radiat. Oncol.
, 63, 1, 27–32, 2019 Feb.
14. Lakhani, P. and Sundaram, B., Deep learning at chest radiography: Automated classification of pulmonary tuberculosis by using convolutional neural networks.
Radiology
, 284, 2, 574–582, 2017 Aug.
15. Ting, D.S., Cheung, C.Y., Lim, G., Tan, G.S., Quang, N.D., Gan, A., Hamzah, H., Garcia-Franco, R., San Yeo, I.Y., Lee, S.Y., Wong, E.Y., Development and validation of a deep learning system for diabetic retinopathy and related eye diseases using retinal images from multiethnic populations with diabetes.
JAMA
, 318, 22, 2211–2223, 2017 Dec 12.
16. Brett, A., Miller, C.G., Hayes, C.W., Krasnow, J., Ozanian, T., Abrams, K., Block, J.E., van Kuijk, C., Development of a clinical workflow tool to enhance the detection of vertebral fractures: Accuracy and precision evaluation.
Spine
, 34, 22, 2437–2443, 2009 Oct 15.
17. Maheshwari, J. and Mhaskar, V.A.,
Essential orthopaedics: (including clinical methods)
, Jaypee Brothers Medical Publishers, New Delhi, India, 2019 Feb 28.
18. Meena, T. and Roy, S., Bone fracture detection using deep supervised learning from radiological images: A paradigm shift.
Diagnostics
, 12, 10, 2420, 2022 Oct 7.
19. Yasaka, K. and Abe, O., Deep learning and artificial intelligence in radiology: Current applications and future directions.
PLoS Med.
, 15, 11, e1002707, 2018 Nov 30.
20. Nakaura, T., Higaki, T., Awai, K., Ikeda, O., Yamashita, Y., A primer for understanding radiology articles about machine learning and deep learning.
Diagn. Interventional Imaging
, 101, 12, 765–70, 2020 Dec 1.
21. Montagnon, E., Cerny, M., Cadrin-Chênevert, A., Hamilton, V., Derennes, T., Ilinca, A., Vandenbroucke-Menu, F., Turcotte, S., Kadoury, S., Tang, A., Deep learning workflow in radiology: A primer.
Insights into Imaging
, 11, 1–5, 2020 Dec.
22. Lodwick, G.S., Keats, T.E., Dorst, J.P., The coding of roentgen images for computer analysis as applied to lung cancer.
Radiology
, 81, 2, 185–200, 1963 Aug.
23. Kalmet, P.H., Sanduleanu, S., Primakov, S., Wu, G., Jochems, A., Refaee, T., Ibrahim, A., Hulst, L.V., Lambin, P., Poeze, M., Deep learning in fracture detection: A narrative review.
Acta Orthop.
, 91, 2, 215–220, 2020 Mar 3.
24. Rainey, C., McConnell, J., Hughes, C., Bond, R., McFadden, S., Artificial intelligence for diagnosis of fractures on plain radiographs: A scoping review of current literature.
Intell.-Based Med.
, 5, 100033, 2021 Jan 1.
25. Urakawa, T., Tanaka, Y., Goto, S., Matsuzawa, H., Watanabe, K., Endo, N., Detecting intertrochanteric hip fractures with orthopedist-level accuracy using a deep convolutional neural network.
Skeletal Radiol.
, 48, 239–244, 2019 Feb.
26. Szegedy, C., Vanhoucke, V., Ioffe, S., Shlens, J., Wojna, Z., Rethinking the inception architecture for computer vision, in:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
, pp. 2818–2826, 2016.
27. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Imagenet large scale visual recognition challenge.
Int. J. Comput. Vision
, 115, 211–252, 2015 Dec.
28. Lindsey, R., Daluiski, A., Chopra, S., Lachapelle, A., Mozer, M., Sicular, S., Hanel, D., Gardner, M., Gupta, A., Hotchkiss, R., Potter, H., Deep neural network improves fracture detection by clinicians.
Proc. Natl. Acad. Sci.
, 115, 45, 11591–11596, 2018 Nov 6.
29. Tomita, N., Cheung, Y.Y., Hassanpour, S., Deep neural networks for automatic detection of osteoporotic vertebral fractures on CT scans.
Comput. Biol. Med.
, 98, 8–15, 2018 Jul 1.
30. He, K., Zhang, X., Ren, S., Sun, J., Deep residual learning for image recognition, in:
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
, pp. 770–778, 2016.
31. Khatik, I., A study of various bone fracture detection techniques.
Int. J. Eng. Comput. Sci.
, 6, 5, 21418–23, 2017 Jun.
32. Hodge, V. and Austin, J., A survey of outlier detection methodologies.
