Using Deep Learning to Improve Early Diagnosis of Pneumonia in Underdeveloped Countries Kyler Larsen

2025-05-06 0 0 233.94KB 11 页 10玖币

侵权投诉

Using Deep Learning to Improve Early Diagnosis

of Pneumonia in Underdeveloped Countries

Kyler Larsen

November 2021

1 Introduction

Increasing eﬃciency and accuracy for the diagnosis of pneumonia can decrease

thousands of preventable deaths each year. Pneumonia is a common disease

in the community that can be easily treated with inexpensive oral antibiotics

administered by volunteer health care workers (Pneumonia 2021). However, if

there is any delay in diagnosis or treatment, the disease course can be very

serious and even deadly. Many of these diagnoses can be made through identiﬁ-

cation of symptoms such as cough, fever, or an increased respiratory rate. Chest

x-ray is an inexpensive but vital component to diagnose pneumonia and to rule

out other diseases. Certain types of pneumonia, especially atypical pneumonia

such as Mycoplasma pneumoniae, can infect a person for weeks with no discern-

able symptoms. These types of pneumonia account for 10 to 40 percent of the

cases of community acquired pneumonia in the United States (Atypical 2021).

With this type of pneumonia it is even more important to use chest x-rays as

an eﬀective tool for accurate diagnosis.

In underdeveloped countries, speciﬁcally in Sub-Saharan Africa and South-

east Asia, the death rate from pneumonia was almost 8-10 times the death rate

in the United States in 2017 (Dadonaite et al 2021). Although many countries

in Sub-Saharan Africa, such as Cameroon, receive medical equipment, including

x-ray machines, through donation services, up to 70% of the equipment is never

used (Piaggio et al 2009). One primary cause of this problem is the lack of doc-

tors and trained healthcare workers. In the Philippines, it was estimated that

85% of Filipino nurses were working overseas in Western countries, leading to

70% of Filipinos dying without medical attention (Finch 2013). In Sub-Saharan

Africa, a critical shortage of doctors and nurses amounting to 2.4 million per-

sonnel in the years 2000-2006, hampered the ability of many health programs

in place to detect and prevent chronic diseases (Naicker et al 2010). These

statistics show that a shortage of trained medical personnel in underdeveloped

countries causes delayed diagnoses or lack of proper treatment, leading to pos-

sible preventable deaths. This study will focus on evaluating a deep learning

model, speciﬁcally over how well it can detect pneumonia in chest x-rays, and

arXiv:2210.05023v1 [eess.IV] 10 Oct 2022

adopting the model to become an aﬀordable and accurate alternative that will

save many lives.

Previous studies have been conducted over the eﬀectiveness of diﬀerent AI

models in limiting pneumonia’s harmful eﬀects. For example, various machine

learning models were evaluated on their accuracy in predicting severe pneumo-

nia, which would allow proper treatment before the harmful eﬀects could take

place. The models evaluated in the study were the Support Vector Machine, Lo-

gistic Regression model, Random Forest Classiﬁer, Naive Bayes, and AdaBoost.

Ultimately, the study concluded that the Random Forest Classiﬁer performed

the best, due to its superior stability (Luo et al 2020). Another study evalu-

ated a deep learning transfer model, VGG-16, as to how well it could identify

pneumonia caused by COVID-19. The study also used chest x-ray analysis to

form the model’s diagnoses, but concluded that further improvements could be

made by adjusting hyperparameters and transfer learning combinations (M. D.

Hasan et al 2021). This study attempts to ﬁll this gap by evaluating a basic

Convolutional Neural Network (CNN) by experimenting with hyperparameters

to achieve maximum metrics.

The purpose of this study was to evaluate the performance of a CNN model

given input in the form of chest x-rays. This study hypothesizes that the chang-

ing hyperparameters of the model can produce results equivalent to a trained

physician. The null hypothesis assumes the model would not improve through

training or would not achieve results similar to a physician. By testing various

hyperparameters and extracting results based on key metrics, conclusions can

be drawn as to how well the model performed (M. D. Hasan et al 2021). The

signiﬁcance of this study is to provide information regarding the feasibility of

using a basic CNN model to diagnose accurately at point of care when there

is shortage of trained healthcare workers. By developing and evaluating such a

model, this study could save lives by improving eﬃciency and accuracy, as well

as saving money and time.

2 Methods

In this study, the ﬁrst step was to organize the data set. The data set used in this

study is from Kaggle (Patel 2021), and features 6432 pre-classiﬁed images. This

dataset includes 3 separate categories: COVID-19 pneumonia, general pneu-

monia, and normal lung x-rays. For this study, only images from the general

pneumonia and normal lung x-rays were used, amounting to 2400 images with

1200 images from each classiﬁcation. The data was then reshaped and aug-

mented to provide a variety of data and to eliminate the need for “perfect”

data. Augmentation included stretching certain pixels or rotating the image,

allowing the creation of extra data as each image could have multiple augmen-

tations done to it. The data was then split into 80% for training data and 20%

for testing data. The training data is the data used by the model to learn its

methods, while the testing data is used to evaluate the model’s accuracy.

A CNN works by receiving input in the form of an image. Then, through

a hierarchical system, the various layers of the CNN create a network that

resembles connected neurons, similar to the human brain. At each layer, an

activation function transforms the results of each layer into speciﬁc output.

The activation function used in this study is the Rectiﬁed Linear Activation

Unit (ReLU), which returns the sum of each node’s input, or 0 if the sum is less

than 0 (Brownlee 2021). The CNN then applies weights and ﬂattens the output

before it is fed to a feed-forward neural network and backpropagation is applied

to every iteration of training. Over a series of these iterations, the model is able

to distinguish between important dominant features and less-valuable low-level

features (Brownlee 2021).

CNNs are valuable for identifying and classifying images, making it the ideal

model for this study. In this study, the CNN was deﬁned as sequential, since it

only takes in one input and produces only one output, allowing for the addition

of multiple layers. An initial layer of 32 ﬁlters with ReLU activation was then

added, with a max pooling size of (2,2) to reduce output size. Then, additional

layers with 64 ﬁlters were added with ReLU activation before the output was

ﬂattened. The amount of added layers is a tested variable, meaning it needs to

be non-speciﬁc. Dropout requirements were also set at 0.5 in order to reduce

overﬁtting, which is when the model memorizes the data set instead of forming

its own predictions. In this case, 50% of the neurons will be ignored during each

iteration, forcing the model to learn new patterns each time (Brownlee 2021).

Finally, the output is formatted and set to print loss function values, accuracy

metrics, speciﬁcity, and sensitivity.

For the testing phase of the model, experimentation was done through chang-

ing the model’s hyperparameters. The ﬁrst parameter experimented with was

the number of extra layers added to the model, allowing more determinations

to be made at the expense of overﬁtting. Other parameters include the number

of epochs, or iterations, made by the model through the data set during each

training cycle, and the set sensitivity of the model, allowing for the maximum

sensitivity while still holding the other metrics constant. For each change to the

parameters, 5 models were trained and the average values for each speciﬁcity,

accuracy, precision, recall, and f1 score were taken to ensure accurate results.

The output displays the accuracy and speciﬁcity of the model after each model

is trained, as well as a graph depicting the trend of the accuracy over each

epoch.

3 Results

The primary focus of the trials in this study is to maximize sensitivity while

keeping speciﬁcity as high as possible. Sensitivity is deﬁned by

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

UsingDeepLearningtoImproveEarlyDiagnosisofPneumoniainUnderdevelopedCountriesKylerLarsenNovember20211IntroductionIncreasingeciencyandaccuracyforthediagnosisofpneumoniacandecreasethousandsofpreventabledeathseachyear.Pneumoniaisacommondiseaseinthecommunitythatcanbeeasilytreatedwithinexpensiveoralantib...

展开>> 收起<<

Using Deep Learning to Improve Early Diagnosis of Pneumonia in Underdeveloped Countries Kyler Larsen.pdf

共11页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Using Deep Learning to Improve Early Diagnosis of Pneumonia in Underdeveloped Countries Kyler Larsen

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: