Abstract
Fetal anomalies are structural defects in a fetus that can lead to a complicated pregnancy, and disabilities later in life. Early detection and intervention are key in the prevention of later disabilities. Conventionally, specialists detect any anomalies in the fetus by physically analyzing the medical images such as ultrasound scans and MRIs. However, the cost of training a qualified radiologist and the general limitations of human beings such fatigue, lack of speed and experience may lead to delayed or erroneous diagnosis, hence delaying intervention.
In the recent years machine learning has been applied in the detection of conditions such as pneumonia and cancer. This research processes the use of convolutional neural network in the detection of fetal anomalies from ultrasound scans. Due to time limitation, this research Focuses on the detection of only one fetal anomaly, Congenital Talipes Equinovarus (CTEV) which is one of the most common musculoskeletal defects that can be corrected by early detection and intervention. The objective of this study is to develop a deep learning model that can analyze ultrasound scans and detect Congenital Talipes Equinovarus.
200 samples of 2-dimensional ultrasound scans were used in the project, The sample size was split into three main sections: training, validation, and testing data. Three implementations of the model were done and compared: a standard CNN model without augmentation with 67.5% accuracy, a CNN model with augmentation with 77% accuracy and a CNN model with transfer learning with 85%. CNN model with transfer learning was selected to implement the model due to is high accuracy.
Table of Contents
Declaration 2
Acknowledgement 3
Abstract 4
Abbreviations 8
CHAPTER ONE: INTRODUCTION
1.1 Background 9
1.2 Problem Statement 10
1.3 Research Objectives 11
1.3.0 General Objective 11
1.3.1 Specific Objectives 11
1.4 Research Questions 11
1.5 Significance of the Research 11
1.6 Scope and Limitation 12
CHAPTER TWO: LITERATURE REVIEW
2.0 Introduction 13
2.1 Machine Learning 13
2.2 Deep Learning 13
2.3 Neural Networks 14
2.4 Convolutional Neural Networks (CNN) 15
2.5 Convolutional Neural Networks in Medicine 17
2.6 Fetal Anomalies 18
2.7 Related Work 19
2.8 Research Gap 22
2.9 Proposed Model 22
2.8.1 Conceptual Framework of the proposed model 22
2.8.2 Architectural Diagram 23
CHAPTER THREE: METHODOLOGY
3.0 Introduction 24
3.1 Research Design 24
3.1.1 Target Population 24
3.1.2 Sample Selection and Sample Size 25
3.1.3 Study Ethics 26
3.2 Data Acquisition and Analysis 26
3.3 Developing Convolutional Neural Network 27
3.3.1 Development Methodology 27
3.3.2 Model Development 27
CHAPTER FOUR: RESULTS AND DISCUSSIONS
4.0 Introduction 29
4.1 Results 29
4.1.1 Standard CNN (Without Augmentation) 29
4.1.2 CNN With Augmentation 31
4.1.3 CNN with Transfer Learning 33
4.2 Discussion 35
CHAPTER FIVE: CONCLUSION AND RECOMMENDATIONS
5.1 Achievements 37
5.2 Limitations 37
5.3 Further Work 37
Appendices 40
Appendix 1: Code Listings 40
Appendix 2: Budget 46
Appendix 3: Schedule 48
Appendix 4: Resources 48
Appendix 5: Plagiarism Check Report 49
List of Figures
Figure 1: Anatomy of Neural Networks 14
Figure 2: CNN layers 16
Figure 3: Foot of a child with clubfoot/CTEV 19
Figure 4: 2D ultrasound scan of a fetus with clubfoot 19
Figure 5: Conceptual Framework 23
Figure 6: CNN architecture 23
Figure 7: Confusion Matrix 26
Figure 8: Model Accuracy (Standard CNN without Augmentation) 30
Figure 9: Model Loss (Standard CNN without Augmentation) 31
Figure 10: Model Accuracy (Standard CNN with Augmentation) 32
Figure 11: Model Accuracy (Standard CNN with Augmentation) 33
Figure 12: Model Accuracy (Standard CNN with Transfer Learning) 34
Figure 13: Model Loss (Standard CNN with Transfer Learning) 35
Figure 14: Gantt Chart 48
List of Tables
Table 1: Sample size 25
Table 2: Standard CNN Metrics 29
Table 3: CNN with Augmentation Metrics 32
Table 4: CNN with Transfer Learning Metrics 34
Table 5: Project Budget 46
Table 6: Project Schedule 47
Abbreviations
CTEV - Congenital Talipes Equinovarus CNN - Convolutional Neural Networks ML - Machine Learning
ReLU - Rectified linear activation function
CT - Computed Tomography
MRI - Magnetic Resonance Imaging PET - Positron Emission Tomography CTG - Cardiotocography
FHR - Fetal Heart Rate
CHAPTER ONE
INTRODUCTION
1.1 Background
Fetal anomaly is defined as a structural or functional anomaly in the fetus that occurs during the pregnancy period and is detected prenatally, during or after birth [ CITATION Wor20 \l 1033 ]. These defects can complicate pregnancy and cause negative consequences to the developing infant. Other terms used in referring to fetal anomalies include congenital anomalies, congenital malformations, congenital disorders, and congenital abnormalities.
(World Health Organization, 2015) reports that approximately 276,000 newborns worldwide die due to congenital anomalies. Among the different types of structural congenital anomalies, musculoskeletal defects are the most common type of congenital anomalies in the sub-Saharan African countries [ CITATION Fen20 \l 1033 ]. Congenital musculoskeletal anomalies occur when a fetus’ bones, muscle or joints do not develop well or wholly, or some structures are disjointed or not aligned well.
This research will be focusing on the use of convolutional neural networks (CNN) in the detection of Congenital talipes equinovarus (CTEV). Congenital talipes equinovarus, is also known as clubfoot. Club foot is a birth defect that affects the foot, and the foot appears to be bent out of shape. This choice is influenced by the fact that congenital talipes equinovarus is the most reported type of musculoskeletal defect. [ CITATION Edw18 \l 1033 ]
Convolutional Neural Networks, also known as ConvNets or CNNs are a type of Neural Networks that are good at image recognition and classification. Some of the applications that Convolutional Neural Networks have been successful at and known for include face recognition, object recognition and self-driving cars.
This study proposes the use of two-dimensional ultrasound scans to train and test the CNN model. An Ultrasound scan, also known as a sonogram, is a medical test that employs the usage of high frequency sound waves to visualize the internal parts of the body. According to AXA global healthcare, an ultrasound can be charged between 600 to 4000 Kenyan Shillings in Kenyan clinics, making it the most accepted and cost-effective type of imaging modality in Kenya. Additionally, ultrasound scans are the most preferred imaging and monitoring techniques all over the world during the pregnancy period because unlike the other medical imaging techniques, ultrasound does not use radiation.[ CITATION Fan17 \l 1033 ]
The use of deep-learning-based medical image analysis in ultrasound analysis not only promises a great support to doctors and specialists in their decisions, but also gives the first assessment of the probability that the fetus has Congenital talipes equinovarus.
1.2 Problem Statement
Congenital Talipes Equinovarus is a congenital anomaly that affects the foot. If ignored or left untreated, congenital talipes equinovarus can restrain movement by making walking hard, painful, or even impossible. However, early detection and treatment are assumed to be the key to preventing late disabilities.
Although radiologists can interpret ultrasounds and detect anomalies early, it takes several years and a huge financial cost to train a competent radiologist. In addition, a study into the usage of ultrasound in an emergency setup has shown that errors or missed diagnoses occurred in anywhere from eight to 10% of cases.
Ultrasounds are an integral part of pregnancy electronic health records and are presently analyzed and interpreted by human radiologists, who are limited by fatigue, speed, and experience.
1.3 Research Objectives
1.3.0 General Objective
To develop and evaluate a prototype deep learning model that can analyze ultrasound scans and detect Congenital Talipes Equinovarus
1.3.1 Specific Objectives
1. To design a deep learning model to detect Congenital Talipes Equinovarus based on ultrasound images.
2. To train the model with data obtained.
3. To evaluate the model for detection of Congenital Talipes Equinovarus
1.4 Research Questions
To ensure that the objectives of the research are attained, the below research questions are key:
● How can we design a model to detect Congenital Talipes Equinovarus?
● How can we train a model to detect Congenital Talipes Equinovarus?
● How can we test and validate the deep learning model to ascertain accuracy of Congenital Talipes Equinovarus detection?
1.5 Significance of the Research
WHO has estimated that 260,000 deaths worldwide (about 7% of all neonatal deaths) resulted from congenital anomalies in the year 2004. Early detection and treatment and monitoring is key to the reduction of such cases.
Although doctors do a phenomenally good job analyzing findings during prenatal clinics, they are prone to fatigue, lack of experience and errors. Apart from the human limitations faced, the cost of training a qualified medical specialist is high. The cost includes and is not limited to, US$ 48, 169 in tertiary education, US$ 6, 865 in secondary education and US $10, 963 in primary education for a single doctor, all adding up to an average of US$ 65, 997 for each doctor. This being the case, this research will play a part in enhancing accuracy and speed in the detection of Congenital Talipes Equinovarus.
1.6 Scope and Limitation
1. Due to time limitation, this research only covers the detection of one fetal anomaly, namely Congenital Talipes Equinovarus.
2. Data used in this study was obtained from online public repositories of fetus ultrasound scans such as Radiopaedia and https://www.fetalultrasound.com/
Login To Comment