Deep Learning Methods for Lung Cancer Detection,

Classification and Prediction – A Review

Ganga V Saji Thasneem Vazim Sumod Sundar
Centre for Artificial Intelligence Centre for Artificial Intelligence Centre for Artificial Intelligence
TKM College of TKM College of TKM College of
Engineering,Kollam Engineering, Kollam Engineering, Kollam
Kerala, India- 691005 Kerala, India-691005 Kerala, India-691005
2021 Fourth International Conference on Microelectronics, Signals & Systems (ICMSS) | 978-1-6654-4885-7/21/$31.00 ©2021 IEEE | DOI: 10.1109/ICMSS53060.2021.9673598

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

Abstract— Lung cancer is the most common cancer that is fatal if contrast and take the shape of a spherical object. The patient survival
treated late. If the disease could be found at an earlier stage before rate can be greatly increased if the lung nodules can be consistently
it’s severity, it is more likely to be treated and diagnosed diagnosed at an early stage. However, disease in lung nodules cannot
successfully. The presence of lung cancers can be detected from be discovered rapidly using raw chest X-rayimaging on a large scale,
computed tomography and chest x-ray images by locating enlarged analysing these diagnostic images manually has become a tedious
lymph nodes. The spread of disease around these nodes can be and difficult task. Deep Learning approaches had made a significant
identified by characterizing size, shape and location; thus, assist
contribution to the progress of medical image analysis by improving
doctors in detecting lung cancers at early stages. In many cases, the
lung cancer diagnosis is based on doctors' experience, which might prediction accuracy, which has resulted in better planning and
lead to misdiagnosis and cause medical issues in patients. There diagnosis. Deep learning- based automated medical diagnosis aids
have been numerous strategies and methods for predicting level of medical experts in the detection and diagnosis of diseases. In this
cancer malignancy using deep learning and machine learning paper, we have discussed different deep learning methods for the
methods.. In this paper, we have studied different Deep Learning prediction, classification and detection of cancerous lung nodules
methods used for the detection, classification and prediction of along withthe identification of their malignancy levels.
cancerous lung nodules and the identification of their malignancy
levels. We have analyzed the advantages and limitations of each Deep neural networks that used for predicting cancerous lung
method along with various datasets used and they are summarized. nodules are analyzed first; models like Adaptive Hierarchical
Heuristic Mathematical Model (AHHMM) are used for lung
Keywords—Convolutional neural network, deep learning, deep cancer prediction from Computed Tomography (CT) images [1].
neural network, lung cancer, machine learning. As the second task is classification, various methods are used in
classifying lung nodules and their malignancy such as Deep
Neural Networks, Convolutional Neural Network and Stacked
Auto Encoder (SAE) [3] are studied next. A Multi View
Knowledge Based Collaborative (MV-KBC) Model
I. INTRODUCTION segmentation is another model which helps to classify lung
Lung cancer is one among the major causes of death in both cancer accurately [4]. The next task that is reviewed is detection
men and women, with an estimated death of five million each based on Improved Profuse Clustering Technique (IPCT)
year and thus a major cause of medical severity worldwide. Lung integrated with Deep Learning Instantaneously Trained Neural
cancer accounts for roughly 1:8 million new cases per year (13 Network (DITNN) model [7] and Deep Learning model with
percent of all tumors) and 1:6 million deaths (19:4 percent of all Wavelet De-noising analysis method[8]. Finally, the methods
tumors) worldwide, with a survival rate of about 18 percent. As using a combination of detection and classification on lung
a result, a lot of medical practitioners and researchers have started nodules such as Automated deep learning 3-Dimensional
to experiment on lung cancer detection and therapy. Lung cancer Customized Mixed Link Network architecture with R-CNN and
is caused by a tumor termed as a nodule, that develops from cells GBM[2] , Densely connected convolution neural network
in the respiratory system's airways.The symptoms cannot (DenseNet) with adaptive boosting algorithm[9] and Detection
beusually identified until the disease has progressed to an and Classification in Whole-slide Lung Histopathology system
advanced stage. In most cases, the symptoms are misinterpreted [6] are studied. In section 2, we are discussing these deep learning
as other issues like an infection or long-term smoking methods along with their algorithms and datasets associated in
consequences. This may delay the diagnosis and treatment of detail. Also, we carried out an exhaustive review of various deep
lung disease. In chest X-rays, these cells are always in direct automated tasks that exhibited top performance in lung cancer

978-1-6654-4885-7/21/$31.00 ©2021 IEEE

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:29:53 UTC from IEEE Xplore. Restrictions apply.
 diagnosis. In Section 3, deep analysis of these techniques is pulmonary malignancies and test the deep learning algorithm for
presented; their performances are inferred along with the lung cancer detection in PET imaging with low dose settings.
challenges and future scopes. Accuracy of the lung cancer detection using deep learning
algorithm on FDGPET (Fluorodeoxyglucose Positron emission
tomography) was evaluated first. Then the performance of the
II. TASK REVIEW algorithm was assessed for simulated lowered level FDG dose
injection. 3936 PET slices were exported, which included scans
2.1. Prediction of lung tumors and image slices of patients with no lung cancer.
The neural network's diagnostic performance is evaluated with
Yu H et al. [1] proposed an Adaptive Hierarchical Heuristic
clinical standard dosage PET images, as well as images
Mathematical Model (AHHMM) for lung cancer prediction from
PET10% and PET3.3% decreased radiation dosage. For
Computed Tomography (CT) images. A deep neural network
(DNN) framework that applies the K-means clustering algorithm PET100%, reduced dose PET10%, and PET3.3 % used for
to directly extract high-level attributes from the data was reconstruction, the area under the curve for lung cancer
designed. Experiments were performed to increase the accuracy detection was 0.989, 0.983, and 0.970, respectively. At standard
in prediction. The modified K-means technique classified images dose and ultralow dose PET3.3%, the neural network had a
into slices that are all part of the same image. while DNN sensitivity of about 95.9% and 91.5 %, respectively, and a
classified pictures into slices of images in similar images.The specificity of 98.1% and 94.2%. According to the research, a
method of AHHMM system predicted the CT scan images with completely automated lung cancer diagnosis at an extremely low
high accuracy, about 90% of the images were identified correctly. effective radiation dose may be achievable on integrating
But it is difficult to detect lung cancer and large cell carcinoma advanced PET detector technology with image analysis and
in the early stages. The model was able to detect an accuracy of machine learning algorithms. The system is deficient in its
96.67 % for the absence or presence of lung cancer in the test specificity in screening efforts of lung cancer.
evaluation phase. However, the work can be enhanced by using
Hybridized Heuristic Mathematical Model for the earlier stage Shakeel PM et al. [7] evaluated the CT scan images of lungs
lung cancer predictions. for detecting lung cancer using an IPCT, i.e., Improved Profuse
Clustering Technique and Deep Instantaneously Trained Neural
Network (DITNN) methods from the Cancer Imaging Archive
2.2. Classification (CIA) dataset. The data set has about 5043 DICOM images
Song Q et al. [3] employed CNN, DNN and stacked divided into 3000 training and 2043 testing images. The image
autoencoder to address the issues of benign and malignant quality was improved by applying the probability distribution
classification. The method takes direct input to minimize the data and cumulative distribution technique to compute a weighted
reconstruction complexities in feature extraction and mean function applied on the pixel. Following the enhancement
classification processes. LIDI-IDRI dataset is used in the model. of the image representation, the affected area was segmented
In total, 4581 lung nodule photos was used in the training, using the calculated pixel similarity value. Taking the similarity
including 2265 cases of benign pulmonary nodules and 2311 of measure into consideration, clusters were formed to extract
malignant pulmonary nodules. A total of 448 images which is spectral related features. These features were then trained and
10% of the training data set was cross-validated. All three
classified to perform detection with about 98.42% accuracy and
network architectures were evaluated with the same data set. The
lowest possible classification error (of about 0.038).
results showed that CNN exhibited more performance than the
other two networks.
Li G et al. [8] introduced deep learning model with a method of
Xie Y et al. [4] 2018 proposed a CT scan-based deep learning wavelet de-noising analysis applied. Wavelet dynamic analysis
model for the benign and malignant nodule classification using technique is used to extract and repair the lung parenchyma
MV-KBC (Multi view knowledge-based collaborative) while excluding interference of noise from outside the lung
technique. The 3-D lung nodule features are learnt by dividing a parenchyma. After that, the pulmonary nodule features of CT
3-D nodule into nine views that are fixed. A KBC sub-model is image are extracted using a CNN optimized using a genetic
then built for each view consisting of three different image algorithm and a traditional CNN. Automatic separation of
patches developed to finetune three trained ResNet networks that
corresponding features of different images is done. When the
define the nodules' characteristics. All the nine KBC sub-models
accuracies of the two methods were compared, it was clear that
were employed together to identify lung nodules using the
strategy of adaptive weighting, allowing start-to-end training of the CNN optimized using genetic algorithm emerged more
the MV-KBC model. To minimize the false-negative rate, the accurate. The pulmonary nodule images are detected and also
penalty loss function is implied with minimum impact on the classified using a CNN optimized by a genetic algorithm, which
model's complete performance. With an AUC of about 95.7 gives direction for pulmonary nodule CT image detection
percent, the model classified lung nodules with 91.6 percent technology. Onur Ozdemira et al. [10] proposed a computer-
accuracy. assisted detection and diagnostic technique for lung cancer
screening using low-dose CT scans that yields relevant
2.3. Detection probability estimates. A comprehensive CADe/CADx system
Schwyzer M et al. [5] proposed a neural network model for that uses low-dose CT scans for diagnosing and identifying lung
identifying the lung cancer in patients from controls having no

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:29:53 UTC from IEEE Xplore. Restrictions apply.
 Prediction
Advantages Limitations
SL No . Reference Year Techniques/
Algorithm
1 [1] 2020 Adaptive Hierarchical 1. K-means clustering 1. Early detection or recognition
Heuristic algorithm improves feature of lung cancer and large cell
Mathematical Model extraction carcinoma is difficult
(AHHMM) 2. The weighted mean
estimated by the
convolutional layer replaced
pixel taking up the cumulative
distribution.
3.Provides high prediction
accuracy.
Subtasks:
Segmentation
Classification

Classification
SL No Reference Year Techniques/ Advantages Limitations
Algorithm
1 [3] 2017 1. CNN showed high 1. Limitations of the data sets
2- Deep Neural performance among the three resulted the neural network to
Network, networks. have small layers.
Convolutional Neural 2. Less accurate.
Network and Stacked
AutoEncoder

2 [4] 2018 Multi View Knowledge 1. Compared to other state 1. Lack of specificity.
Based Collaborative of the art methods , it is 2. High computational complexity
(MV-KBC) Model more accurate
2. High accuracy and
sensitivity.
Subtasks:
Segmentation

cancer was used. The model 3D CNNs cascaded and was able classification outcomes are assessed based on parameters such as
to achieve state of the art performance on the detection of lung the patient's age, smoking habits and history, family history,
nodules and level of malignancy on the kaggle datasets and biomarkers, nodule location and size. When integrated with
LUNA16. The work used Monte Carlo dropout and deep clinical considerations, this deep learning model for detecting
ensembles to characterize model’s uncertainty and demonstrated and classifying nodules, aids in the minimizing misclassification
that quantifying model uncertainty allows the system to deliver and false positive (FP) outcomes in initial stages of cancer
calibrated classification probabilities. The calibrated diagnosis. The LIDC-IDRI dataset was used by the model and
probabilities, informed by model uncertainty, can be used obtained 94% sensitivity and 91% specificity.
tomake further diagnostic decisions. However, a patient referral
(or reject option) can be incorporated that could reject the N Kalaivani et al. [9] proposed a DenseNet and adaptive
boosting algorithm for classifying the lung images to normal or
decisions that are uncertain.
malignant. The dataset consists of 201 images, of which 85%
2.4. Detection and Classification ofthe total images were used for training. 15% of the total images
Nasrullah N et al. [2] developed a model for the detection and were used for testing and classification. The method obtained an
classification of malignant nodules in lung by using multiple experimental accuracy of 90.85%.
strategies of deep convolutional neural network. For image Zhang Li et al. [6] proposed an ACDC@ Lung HP system for the
analysis. The model has two 3-Dimensional Customized Mixed evaluation of different diagnosis methods that are computer-
Link Network (CMixNet) architectures for detection and aided for the diagnosing lung cancer. The study employed a
classification of the lung nodules. Faster R-CNN are used upon dataset of 150 training images and 50 test images of 200 patients
efficiently learned features from network encoder and decoder to study cancer tissue segmentation (pixel-wise detection) in
architecture like U-Net to detect nodules. The nodules are WSI. This report examines the challenge and highlights the top
classified using a gradient boosting machine (GBM) based on the ten lung cancer segmentation algorithms submitted. However,
features from the proposed 3D CMixNet framework. Finally, the

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:29:53 UTC from IEEE Xplore. Restrictions apply.
 Detection
SL No Reference Year Techniques/ Advantages Limitations
Algorithm
1 [5] 2018 Neural network model 1. A completely 1. Specificity deficient
with deep learning automated lung
algorithm for lung cancer diagnosis at
cancer detection in extremely low
PET imaging with low effective radiation
dose. dose achieved

2 [7] 2019 Improved Profuse

Clustering Technique 1. Minimum classification 1.High computational cost
(IPCT) and DITNN error
methods 2. The DITNN technique
successfully captures
lung cancer associated
feature as it collects
images throughout the
training.
Subtasks: 3. The effective use of each
Segmentation layer, as well as the
Classification weight update process,
enhances the total
precision value.

Detection and Classification

SL No. Reference Year Techniques/ Advantages Limitations
Algorithm
1 [2] 2019 Automated deep 1. Minimizes 1. Training time is not considered
learning 3-Dimensional misdiagnosis and false while choosing the best model.
Customized Mixed Link positive outcomes in
Network architecture early-stage diagnosis
with R-CNN and GBM. of lung cancer

2. Low computational cost.

2 [9] 2020 A DenseNet with an 1. The method obtained an 1. Require more training time.
adaptive boosting accuracy of 90.85%.
algorithm

3 [6] 2020 Automatic Cancer 1. The work highlights the 1. It does not include the
Detection and top ten lung cancer classification of the primary lung
Classification in Whole- segmentation algorithms cancer subtypes.
slide Lung
Histopathology system

the work does not include the classification of the primary lung
cancer subtypes. lung nodule identification, classification and malignancy
prediction. The capacity of deep learning to outperform other
III. CONCLUSION machine learning algorithms is its most significant advantage.
Lung cancer symptoms cannot be usually identified until the This paper summarizes a detailed analysis on various approaches
disease has progressed to an advanced stage. In most cases, the to detect, classify and diagnose lung malignancies using either
symptoms are mistaken as other issues like an infection or long- CT scan images or X-ray images. Based on the results observed
term smoking consequences. This may delay the diagnosis and from this extensive review, it can be stated that approaches that
treatment of the disease. So, it’s advantageous to detect lung used deep learning techniques produced more accurate results
cancer at an early stage since therapy can then be commenced to than methods that used traditional machine learning techniques.
prevent the disease from becoming detrimental. As a result, this Deep learning neural networks can be used to predict cancerous
work offers a comprehensive review of various deep learning lung nodules using models like AHHMM or lung cancer
algorithms for classifying lung cancers using CT scans or X-ray prediction from CT image [1] was studied. As the second task is
pictures. Deep learning is an emerging and prominent method for classification, various methods used in classifying lung nodules
learning features and recognizing patterns. It serves as an aid in and their malignancy such as DNN, CNN and Stacked Auto
Encoder (SAE) [3] were studied. An MV-KBC Model

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:29:53 UTC from IEEE Xplore. Restrictions apply.
 Segmentation was studied as another model which helps to [4] Xie Y, Xia Y, Zhang J, Song Y, Feng D, Fulham M, Cai W. Knowledge-
based collaborative deep learning for benign-malignant lung nodule
classify lung cancer accurately [4]. Detection based on classification on chest CT. IEEE transactions on medical imaging. 2018
Improved Profuse Clustering Technique integrated DITNN) Oct 17;38(4):991-1004.
model [7] and Deep Learning model with Wavelet De-noising [5] Schwyzer M, Ferraro DA, Muehlematter UJ, Curioni-Fontecedro A,
analysis method [8] were analyzed. Finally, the methods using a Huellner MW, Von Schulthess GK, Kaufmann PA, Burger IA, Messerli
combination of detection and classification on lung nodules such M. Automated detection of lung cancer at ultralow dose PET/CT by deep
neural networks–initial results. Lung Cancer. 2018 Dec 1; 126:170-3.
as automated deep learning 3-Dimensional Customized Mixed
[6] Li Z, Zhang J, Tan T, Teng X, Sun X, Zhao H, Liu L, Xiao Y, Lee B, Li
Link Network architecture with R-CNN and GBM [2], Dense Y, Zhang Q. Deep learning methods for lung cancer segmentation in
Net with adaptive boosting algorithm [9] and detection along whole-slide histopathology images-the acdc@ lunghp challenge 2019.
classification in whole-slide Lung Histopathology system [6] IEEE Journal of Biomedical and Health Informatics. 2020 Nov 20.
were studied. Certain systems lack enough accuracy, while some [7] Shakeel PM, Burhanuddin MA, Desa MI. Lung cancer detection from CT
systems must be developed in order to achieve the maximum image using improved profuse clustering and deep learning
instantaneously trained neural networks. Measurement. 2019 Oct 1;
accuracy level possible. 145:702-12.
[8] Li G, Zhou W, Chen W, Sun F, Fu Y, Gong F, Zhang H. Study on the
REFERENCES detection of pulmonary nodules in CT images based on deep learning.
IEEE Access. 2020 Mar 30; 8:67300-9.
[1] Yu H, Zhou Z, Wang Q. Deep Learning Assisted Predict of Lung Cancer [9] Kalaivani N, Manimaran N, Sophia S, Devi DD. Deep Learning Based
on Computed Tomography Images Using the Adaptive Hierarchical Lung Cancer Detection and Classification. InIOP Conference Series:
Heuristic Mathematical Model. IEEE Access. 2020 May 5; 8:86400-10. Materials Science and Engineering 2020 Dec 1 (Vol. 994, No. 1, p.
[2] Nasrullah N, Sang J, Alam MS, Mateen M, Cai B, Hu H. Automated lung 012026). IOP Publishing.
nodule detection and classification using deep learning combined with [10] Ozdemir O, Russell RL, Berlin AA. A 3D probabilistic deep learning
multiple strategies. Sensors. 2019 Jan;19(17):3722 system for detection and diagnosis of lung cancer using low-dose CT
[3] Song Q, Zhao L, Luo X, Dou X. Using deep learning for classification of scans. IEEE transactions on medical imaging. 2019 Oct 29;39(5):1419-29.
lung nodules on computed tomography images. Journal of healthcare
engineering. 2017 Aug 9;2017.

Authorized licensed use limited to: SARDAR VALLABHBHAI NATIONAL INSTITUTE OF TECH. Downloaded on November 17,2024 at 09:29:53 UTC from IEEE Xplore. Restrictions apply.