Advanced Medical Images Recognition and Diagnosis of

Respiratory System Viruses

Afsana Nur Meem1, Efte Kharul Islam Shinha1, Iffat Jahan Shorna1 ,Lamya Ishrat Nodi1,
*Ahmed Wasif Reza1
1
Department of Computer Science and Engineering, East West University,
Dhaka, Bangladesh

Emails: [email protected], [email protected],

[email protected], [email protected], [email protected],

Abstract Diagnostic and treatment monitoring of lung cancer at an initial point

requires the use of medical imaging techniques. Lung cancer detection has been
thoroughly researched using a variety of medical imaging modalities, including
computed tomography, positron emission tomography, magnetic resonance imaging,
chest X-ray, and molecular imaging methods. These strategies have several
restrictions, such as not immediately categorizing cancer photos, making it
inappropriate for patients with other illnesses. A sensitive and precise method for the
early diagnosis of lung cancer must be developed immediately. With quickly
developing applications across medical image-based and textural data modalities,
deep learning is one of the fastest-growing subjects in medical imaging. Clinicians
can detect and categorize lung nodules more quickly and efficiently with the aid of
deep learning-based medical imaging systems. The most current advancements in
imaging methods based on deep learning for the early diagnosis of lung cancer are
discussed in this study.

Keywords: lung cancer, medical images, segmentation, classification, deep learning,

convolutional neural network

1. Introduction

Cancer with the highest morbidity and fatality rates in the US is lung cancer, which is also
the most common disease [1]. GLOBOCAN estimated that there were 1.76 million lung
cancer-related fatalities and 2.09 million new cases in 2018 [2]. Globally, there has been a
considerable increase in lung cancer diagnoses and fatalities [2]. Non-small cell lung
carcinoma (NSCLS) accounts for about 85-88% of lung cancer cases, while small cell
lung cancer (SCLC) accounts for about 12-15% of lung cancer cases [3].

1
Due to lung cancer's invasiveness and heterogeneity, early identification and treatments
are essential to boosting the disease's overall 5-year survival rate [4].

The detection of lung nodules has been thoroughly studied over the past 20 years using a
variety of medical imaging techniques, including chest X-ray, positron emission
tomography (PET), magnetic resonance imaging (MRI), computed tomography (CT),
low-dose CT (LDCT), and chest radiograph (CRG). Although CT is the gold standard
imaging method for finding lung nodules, it has a high false-positive rate, can only find
apparent lung cancer and emits dangerous X-ray radiation [5].

For the purpose of detecting lung cancer, LDCT has been proposed [6]. However,
participants undergoing LDCT accounted for the majority of cancer-related mortality. The
development of 2-deoxy-18F-fluorodeoxyglucose (18F-FDG) PET has improved the
ability to identify lung cancer [7]. When used to diagnose NSCLC, 18F-FDGPET
generates semi-quantitative characteristics of tumor glucose metabolism [8]. However,
patients with NSCLC require additional assessment because of 18F-FDG PET.

2. Lung Imaging Techniques

Radiologists can diagnose lung disease with the aid of medical imaging technologies. In
comparison to these medical imaging techniques, CT has greater benefits, including
information on lesion growth, size, location, characterization, and nodule information,
which could help identify lung cancer. The management of lung cancer is considerably
impacted by the use of 4D CT because it allows for more precise radiation targeting [21].
Using an optimal deep neural network (ODNN) and linear discriminate analysis (LDA),
In this paper [9], created an autonomous detection method to categorize lung cancer in CT
lung pictures. For the purpose of reducing feature dimension, the LDA lowered the
extracted picture features. A modified gravity search technique was used to apply the
ODNN and optimize it, producing a classification result that is more accurate.
LDCT uses less radiation than CT and is more sensitive to the early detection of lung
nodules and cancer. The death rate from lung cancer is not, however, decreased. For high-
risk smokers aged 55 to 74, it is advised that LDCT be done once a year [10].
The most popular unsupervised learning algorithms in medical pictures are convolutional
neural networks (CNN), deep convolutional neural networks (DCNN), and recurrent
neural networks (RNN). Because minimal preprocessing is needed, CNN architecture is
one of the most used supervised deep learning methods for lesion segmentation and
classification. Recently, CNN architectures have been used to classify and segment
medical pictures.

3. Literature Review

2
Lung nodule identification is difficult since their shape, texture, and size vary widely, and
certain non-nodules, including blood vessels and fibrosis, closely resemble lung nodules
that frequently arise in the lungs. The procedure consists of two major steps: lung nodule
identification and false-positive nodule reduction.

This paper [11] looked into how lung nodule identification in chest CT images was
affected by the efficacy of deep learning image reconstruction (DLIR) techniques. This
study placed up to six artificial nodules into the lung phantom. Images were created with
50% ASIR-V and DLIR strengths of low (DL-L), medium (DL-M), and high (DL-H).
There was no statistically significant difference between these techniques.

This paper [12] evaluated the correlation between CT and commercial CAD for the
detection of lung nodules. They also examined LDCT pictures from 36 human individuals
using three reconstruction kernels (B, C, and L). All images had intensities of 82%, 88%,
and 82% for the nodules of B, C, and L, accordingly. Experiment results revealed that
merging data from two distinct kernels without exposure to radiation could increase CAD
sensitivity.

Based on a quicker R-CNN network and an FP reduction model in thoracic MR images,

[13] created a method for detecting lung nodules. A quicker R-CNN was constructed in
this study to detect lung nodules, and an FP reduction model was developed to reduce FP.
The approach was evaluated on the FAHGMU dataset and yielded an 85.2% sensitivity
with 3.47 FP per scan.

This paper [14] investigated the impact of lowering the CT dose on the effectiveness of a
CAD-based nodule detection model. The CAD system was tested on the NLST dataset
and yielded sensitivities of 35%, 20%, and 42.5% at the initial dose, 50% dose, and 25%
dose, accordingly.

For the purpose of identifying lung nodules, this paper [15] created the two-stage CNN
(TSCNN) model. A ResDense-based U-Net was used in the first stage to identify lung
nodules. In the second step, a 3D CNN-based ensemble learning architecture was
suggested to minimize false-positive nodules. Three existing models were compared to the
proposed model, including 3DDP-DenseNet, 3DDP-SeResNet, and 3DMBInceptionNet.

Papers Methods Datasets Advantages Dis- Results Research Gap

advantages

[11] They have In this The growth It may be They found We can observe
used deep study of multiplex more out how lung from this research
learning images PCR tests for expensive for nodule that to assess the

3
 image were the detection the laboratory identification effects of the
reconstructio created of multiple to detection in chest CT medications and
n (DLIR) with 50% respiratory of respiratory images was help clinicians
techniques. ASIR-V viruses opens viruses. affected manage their
and DLIR the way for patients, new
strengths future diagnostic tests will
multiplex
of low be needed.
tests for the
(DL-L),
detection of
medium viruses and
(DL-M), other groups
and high of viruses
(DL-H) present in
clinical
specimens.

[12] They have They have It can be They have The classifier Experiment results
used the examined used to used a few they used in revealed that
correlation LDCT identify data to detect their research merging data from
between CT pictures diseases lung nodules demonstrated two distinct kernels
and from 36 easily and out of CT good accuracy without exposure to
commercial human rapidly with images. which can radiation could
CAD for accuracy. detect out of
individual increase CAD
detection. CT images.
s using sensitivity.
three
reconstruc
tion
kernels

[13] They have The The Their method Their method, They have used a
used the R- approach approach only works on R-CNN very small number
CNN was was the moderate network and of data, large data
network and evaluated evaluated dataset. an FP can differ in the
an FP on the and yielded reduction accuracy rate and
reduction FAHGMU an 85.2% model in prediction.
model for thoracic MR
dataset. sensitivity
detecting images which
with 3.47 FP
lung nodules can be useful
per scan. for lung
nodules
detection.

[14] They have The CAD The finding They only Based on their They should have

4
 examined the system of this will used CAD- findings, used some other
impact of was tested assist based sensitivities of image-processing
lowering the on the clinicians in detection to 35%, 20%, techniques for the
CT dose on NLST improving examine. and 42.5% at detection.
the dataset patient the initial
effectiveness management dose, 50%
of a CAD- methods. dose, and
based nodule 25% dose,
detection accordingly.
model.

[15] They have They Their It is a lengthy In their first For detecting the
created the compared methods process to stage they they should have
two-stage three were specific detect lung identify lung used other easy
CNN existing for nodules by nodules. In image-processing
(TSCNN) models to identifying using this the second techniques.
model. A the lung nodules method. step, a 3D
ResDense- detection and CNN-based
proposed
based U-Net TSCNN ensemble
model,
was used in model. learning
the first stage including architecture
to identify 3DDP- was used to
lung nodules. DenseNet, minimize
In the second 3DDP- false-positive
step, a 3D SeResNet, nodules.
CNN-based and
ensemble 3DMBInc
learning eptionNet.
architecture
was

4. Materials and Methods

This section presents recent lung cancer and nodule prediction achievements using deep
learning techniques. The processing includes image pre-processing, lung nodule
segmentation, detection, and classification.

a. Imaging Pre-Processing Techniques and Evaluation

b. Pre-Processing Techniques

5
Recent developments in lung cancer segmentation, detection, and classification utilizing
deep learning techniques are discussed in this work. The most cutting-edge deep learning-
based lung cancer detection techniques are highlighted in this article. This essay also
emphasizes current successes, pertinent research roadblocks, and upcoming research
directions. The remainder of the essay is organized as follows. Section 2 describes the
currently available medical lung imaging techniques for lung cancer detection, Section 3
reviews some recently developed deep learning-based imaging techniques, Section 4
presents deep learning-based lung cancer prediction, Section 5 discusses the current
challenges and future directions of deep learning-based lung imaging methods, and
Section 6 concludes this study.

Bengin Cases

Malignant Cases

6
Normal Cases

Model Accuracy

7
5. Conclusion
This paper proposed a simple Convolutional neural network for image classification. This
simple convolutional neural network imposes a less computational cost. On the basis of
the convolutional neural network, we also analyzed different methods of learning rate sets
and different optimization algorithms for solving the optimal parameters of the influence
on image classification. We also verify that the shallow network also has a relatively good
recognition effect.

References

[1] J. Kuruvilla and K. Gunavathi, ”Lung cancer classification using neural networks for
CT images”, Computer Methods and Programs in Biomedicine, vol. 113, no. 1, pp.
202-209, 2014.

[2] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, Gradient-based learning applied to

document recognition, Proceedings of the IEEE, vol. 86, no. 11, pp. 22782324,
1998.

[3] Ciregan,D.,U.Meier, and J. Schmidhuber. Multi-column deep neural networks for

image classification. in Computer Vision and Pattern Recognition (CVPR), 2012

8
 IEEE Conference on. 2012. IEEE..

[4] Ian J. Goodfellow, D.W.-F., Mehdi Mirza,Aaron Courville,Yoshua Bengio, Maxout

networks. 2013. 28: p. 1319-1327

[5] Hu, M.L.a.X., Recurrent Convolutionalal neural network for object recognition.
2015.

[6] Scott rozelle, Xue Lei Xu Yangming, etc. The depth of the study review [J].
Computer application research, 2012, 29 (8) : 2806-2810.

[7] K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale
image recognition,” in ICLR, 2015.

[8] Otsu, N. A threshold selection method from gray-level histograms. IEEE Trans. Sys.
Man Cyber. 1979, 9, 62–66.

[9] Hu, S.; Hoffman, E.A.; Reinhardt, J.M. Accurate lung segmentation for accurate
quantization of volumetric X-ray CT images. IEEE Trans. Med. Imaging 2001, 20,
490–498

[10] Wang, S.; Zhou, M.; Olivier, G.; Tang, Z.C.; Dong, D.; Liu, Z.Y.; Tian, J. A Multi-
view Deep Convolutional Neural Networks for Lung Nodule Segmentation. In
Proceedings of the 2017 39th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (EMBC), Jeju, Korea, 11–15 July
2017; pp. 1752–1755.

[11] Li, Y.; Zhang, L.; Chen, H.; Yang, N. Lung nodule detection with deep learning in
3D thoracic MR images. IEEE Access 2019, 7, 37822–37832.

[12] Khosravan, N.; Bagci, U. Semi-Supervised Multi-Task Learning for Lung Cancer
Diagnosis. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. 2018, 2018, 710–713.

[13] Jiang, J.; Hu, Y.C.; Liu, C.J.; Darragh, H.; Hellmann, M.D.; Deasy, J.O.; Mageras,
G.; Veeraraghavan, H. Multiple resolution residually connected feature streams for
automatic lung tumor segmentation from CT images. IEEE Trans. Med. Imaging
2019, 38, 134–144

[14] Zhang, M.; Li, H.; Pan, S.; Lyu, J.; Su, S. Convolutional neural networks based lung
nodule classification: A surrogate-assisted evolutionary algorithm for
hyperparameter optimization. IEEE Trans. Evol. Comput. 2021, 25, 869–882

[15] Cao, H.C.; Liu, H.; Song, E.; Ma, G.Z.; Xu, X.Y.; Jin, R.C.; Liu, T.Y.; Hung, C.C. A

9
Two-Stage Convolutional Neural Networks for Lung Nodule Detection. IEEE J.
Biomed. Health Inform. 2020, 24, 2006–2015

10