2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE
Lung Tumor Classification and Detection from CT
Scan Images using Deep Convolutional Neural
Networks (DCNN)
N. Mohanapriya, B. Kalaavathi,
VCEW, Tiruchengode, K.S.R IET,
India.
India
mohananatesan@gmail.com
Kalabhuvanesh@Yahoo.Com
Abstract - Lung cancer is life-threatening diseases
and it now affects all people regardless of gender. Precise lung
tumor classification helps to diagnosis lung cancer early, which
decreases the rate of death of lung cancer, but it is hard to
detect early. Harmless or malignant is known as a lung tumor.
It is innocuous when the tumor cells were healthy, when the
cells are abnormal and can grow hysterically, they are
cancerous cells, and the tumor is in curable. A classifier based
on Deep Convolutional Neural Networks (DCNN), which
classifies the lung tumor composed of different fully connected
pooling and Convolutional layers. Three architectures were
defined for DCNN classifier each one is trained with different
patch size. DCNN is applied to the CT image for classification
of benign and malignant lung tumor. The proposed
architectures were examined on the LIDC database and cross
checked with other classifiers result such as Artificial Neural
Network Simulation result presents DCNN classifier achieves
better performance.
Keywords: Deep Learning, Convolutional Neural
Network, Tumor, CT images, Benign and Malignant.
I. INTRODUCTION
Recent statistics from the world health institutions
demonstrate that approximately 7.6 million people die every
year in the world due to cancer. In regards, it is expected
which cancer mortality will continue to rise to around 18
million worldwide in 2030. The main causes of death are
Lung cancer mortality rate is counting for about 27% of all
deaths from cancer [2]. The death of the lung cancer patients’
numbers is more-than that of breast cancer. Due to lack of
literacy towards lung cancer, the possibility of detecting the
disease is harder at the early stage. In other words, at the
time of detection, already the disease reach to its advance
stage [3]. The stages of lung cancer can be classified as first
and second stage. The First stage refers to tumor presents
only in limited area and the second stage is tumors that
expand over to various body parts. Deep learning algorithms,
especially DCNN, have rapidly become a popular
methodology for processing medical images as well,
including detection of medical problems, Feature Extraction,
segmentation, and classification.
978-1-7281-3778-0/19/$31.00 ©2019 IEEE
800
T. senthil Kuamr
HCL Technologies,
Chennai, India
thanapals@hcl.com
This paper proposes a DCNN based architecture for
classifying the lung tumor. If the tumor is benign then chances
of cancer could be low and cancer can be initial stage and also
curable. When the tumor becomes malignant then chances of
cancer to be high. Deep neural networks consisting of multiple
hidden layers such as convolutional layer, pooling layer, and
fully connected layer. Convolutional layer is characterized by
local connectivity and weight sharing. This layer's nodes are
grouped into feature maps that share the same weights, weight
sharing radically reduces the number of network parameters and
thus increases efficiency and prevents over fitting. Through
aggregating small sub-sets of values, the pooling layer is used to
subsample the previous layer. Max or average pooling
substitutes input values by maximum or average values to
reduce output sensitivity to minor input changes. The
classification results are generated by a fully connected layer.
II. RELATED WORK
An image segmentation using ANN for detecting the
lung cancer using X-Ray images. Artificial neuron consists of
three basic sets of rules, multiplication, summation, and
activation. Initially, Seed fill operation is applied; it performs the
process on binary image background pixels starting from the
points. This reveals the images that struggle with 60%, 70% and
80% of lung cancer. Lung Cancer Prediction [5] is performed
using the genomic features of the Oriented Gradient Histogram
(HOG) all together with Artificial Neural Network. Different
phases were included for implementing the prediction model.
Such as pre-processing based on data acquisition and
classification based on extracted features. After feature
extraction, the HOG descriptor is investigated to remove agent
highlights from the images of the genomic arrangements from
Voss encoder. To choose the hyperbolic digression sigmoid
capacity for the yield layer neurons because it has the capability
to handle either ceaseless qualities or -class binary
arrangement issues.
To upgrade the structures of the fragmented images,
decrease the bogus identification and abstain from missing
knobs joined on the lung walls. After this, intensity based
highlights are determined from every knob. The authors of [7]
segment the complete lung and classified from the CT scans
picture using neural networks. The arguments are determined
from the sectioned image using the standard deviation, mean
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE
statistical arguments, kurtosis, and skewedness. The process
of classification is accomplished through feed forward and
feedback neural networks. The feed-forward propagation
system provides better classification contrasted with feed-
forward systems..
In designed the CAD framework for lung nodules
classification with artificial neural networks. It gives
division of knobs on the projections with neural systems
grouping among favorable and dangerous knobs with the
assistance of ANN. It also provided better accuracy,
sensitivity and specificity and it used an aggregate of 128 CT
scans acquired from 47 patients. ANN uses three sections in
the creator section: input, hidden and output sections. 22
neurons in just one hidden layer were used in the
investigation. A tale aspiratory knob recognition strategy
dependent on Progressive Square utilizing Support Vector
Machine (SVM) grouping is proposed in [9] and it comprises
of three stages. The initial step, split information CT pictures
into square pictures and non-enlightening squares are sifted
through. In the subsequent advance, knob applicants are
identified. Fragment the articles in the chose square pictures
and change their area to identify the knob. Finally classify
the nodule applicants into knobs and non-knobs. Highlight
vectors are removed from the item in the knob applicant
squares. SVM classifiers are applied to the removed element
vectors to recognize knobs. This methodology has decreased
the bogus encouraging points in the knob fundamentally and
it accomplished 95.28% affectability.
An algorithm which efficiently identifies and
classifies tumor in the CT images. Initially the enhancement
technique is applied by means of Wavelet based Adaptive
Histogram Equalization (WAHE) to correct the contrast
which in turn increases the accuracy of lung tumor
segmentation. Next in the segmentation phase, Level set
based Cellular Automata (LCA) was applied to extract the
tumor region from CT images efficiently. A neural network
based approach presented to detect lung cancer from raw
chest images. These extracted highlights are assumed as the
inputs the neural network. Further, it is verified whether the
extracted nodule is a non-malignant or malignant. The
training dataset of CT images are processed in three stages to
attain more quality and accuracy in the processed
examination. For segmentation purpose Fuzzy C-Means
(FCM) technique is used. The three kinds of deep neural
networks are Stacked Auto encoder (SAE). The trial results
show that the CNN system filed the better execution with an
exactness, affectability and explicitness which have the best
outcome among the three systems. Computer Aided
Detection (CAD) framework for feature portrayal and
classification assume basic jobs in false-positive reduction
(FPR). Further, structure a profound convolutional neural
systems strategy for knob arrangement.
III. PROPOSED WORK:
The presence of pulmonary tumor, solid clumps of
tissue that appear in and around the lungs, is a key feature of
lung cancer. These tumors are visible in CT scan images and
can be malignant (cancerous) in nature, or benign (not
801
cancerous). The shape and size of tumors vary, and benign
tumors can look very similar to malignant ones. Furthermore,
the presence of blood vessels in the lung makes distinguishing
tumors a challenging task, especially on 2D image slices.
A. Introduction to Proposed Classifier:
The proposed Classifier is divided into four separate
stages. They are the: (1) Tumor detector, which accepts a CT
scan and identifies areas that contain tumor; (2) malignancy
detector, which operates similarly to the tumor detector, but
further classes tumor as benign (non-cancerous) or malignant
(cancerous); (3) tumor classier, which accepts individual tumor
volumes and similarly classes them as benign or malignant; and
(4) patient classifier, which accepts features from the
malignancy detector and tumor classifier and yields the
probability of the patient having tumor.
The tumor detector is used to detect regions that
contain tumor. A malignancy detector provides a class
probability map over grid cells of each cell containing benign,
malignant, or normal (no tumor). Separate code extracts and
preprocesses tumor volumes and runs the classier on each,
yielding the probability of malignancy for each tumor. Finally,
the patient classifier pools features from the tumor classifier and
the malignancy detector and produces the probability of the
patient having lung tumor. Figure 1 illustrates the Lung Tumor
Classification Process.
1) Preprocessing:
Noise in medical images is caused by different sources,
including many external causes in the transmission system and
environmental facts. The noise removal methods have become
an essential factor for accurate detection of lung tumor.
2) Deep Convolutional Neural Networks Architecture:
The Convolutional Neural Networks model by
transforming input image with a linear filter, adding a bias term
and applying a non-linear function, this model can learn lots of
feature maps. Three different Convolutional Neural Networks
architecture each with three different patch sizes was proposed
here to classify the tumors. A detailed result of Convolutional
Neural Networks with input patch 16 × 16 size of six layers is
illustrated in Figure 2, this work included 2 convolutional
layers,2 max-pooling layers, a fully-connected layer and a
softmax classifier.
The convolutional layer C1 ultimate layer contained
two units completely associated with the layer F6, one
neuron actuated by softmax relapse created tested with three
different patch sizes 16×16, 24×24 and 32×32. The results of
these models were presented in the next section. As the
complexity of the network increase, the execution time also
increases. The key challenge is to balance the accuracy and the
execution time. We use the limited feature map, if we increase
the number of feature maps the accuracy will be improved.
These Convolutional Neural Networks architectures are
illustrated in Table 1 was used to implement the Convolutional
Neural Networks model.
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE

CT Lung Image NO Normal

Preprocessing DCNN
Image
Classifier

YES
Tumor Detected

Benign Malignant

Fig..1. Lung Tumor Classification Process

Softmax Classifier
C2: P2: 6*6
Patch: 16×16 C1:32@14*1 32@12*12
P1: 7*7 Normal
4

Benign

CT Lung Image Maligna

nt
5×5 2×2 3×3 2×2

Figure 2. Proposed DCNN Architecture

TABLE 1. THREE DCNN ARCHITECTURES

Architecture / Architecture 1 Architecture 2 Architecture 3
Layer
C1 : 32 Feature Map, Size of 5 *5 and 1 C1 : 32 Feature Map, Size of 5 *5 C1 : 32 Feature Map, Size of 5 *5 and 1
Layer 1
Stride Convolutional Layer and 1 Stride Convolutional Layer Stride Convolutional Layer

P1: Max Pool of 2*2 and stride 2 (with P1:Avg Pool with 2×2 and stride 2 P1:Max Pool with 2×2 and stride 2 (with
Layer 2
padding) (without padding) padding)

C2 : 32 Feature Map, Size of 3×3 and C2 : 32 Feature Map, Size of 3×3 C2 : 32 Feature Map, Size of 3×3 and 1
Layer 3
1 Stride Convolutional Layer and 1 Stride Convolutional Layer Stride Convolutional Layer

Layer 4 P2: Max Pool of 2*2 and stride 2 P2: Avg Pool with 2×2 and stride 2 P2: 2×2 and stride of 2 Max Pool

C3 : 64 Feature Map, Size of 3×3 and C3 : 64 Feature Map, Size of 3×3 and 1
Layer 5 FullyConnected Layer
1 Stride Convolutional Layer Stride Convolutional Layer

Layer 6 Softmax Classifier Fully Connected Layer Fully Connected Layer

Layer 7 - Softmax Classifier Softmax Classifier

802
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE

IV. RESULT AND DISCUSSION

The lung tumor detection and classification problem
is considered here and which is resolved using DCNN.
The Table 2 shows the accuracy of the DCNN architectures.
The proposed work accepts the CT scanned images as Each architecture composed of different number of layer and
input which is accessed from LIDC datasets. and it is evaluated with three patch sizes (16×16), (24×24),
(32×32). The architecture 3 yields the higher accuracy 24 × 24
Results and Parameters Evaluation patch size it is evaluated with three patch sizes (16×16),
The parameters used for evaluating the performance (24×24), (32×32).
of proposed DCNN architecture are Precision, Recall, Figure 3 shows the graphical representation of accuracy of the
Dice Score Coefficient and Accuracy. proposed DCNN architectures. Each architecture composed of
1.1 Precision: different number of layer and it is evaluated with three patch
Precision is used to measure how relevant search sizes (16×16), (24×24), (32×32).
results are in following equation:
Precision = Positive/ True positive+ false positive (1)
1.2 Recall:
Recall is used for how complete the results are and it
is calculated with the following equation,

1.3 Dice Similarity Co-efficient (DSC):

Dice Similarity is used to measure how similar the
objects are. Hence the size of overlap of the two
segmentation by the total size of the two objects.

1.4 Accuracy: Fig. 3. Accuracy of Proposed DCNN Architectures

The proposed DCNN architectures are compared and
performance of each architecture evaluated by accuracy.
Accuracy of a measured value to an actual value and also
precision is independent of accuracy, which calculated by,
The architecture 3 yields the higher accuracy 24 × 24 patch
size
TABLE 2. COMPARISON OF ACCURACY FOR PROPOSED DCNN ARCHITECTURE
Table 3 results comparison of proposed DCNN and the
existing techniques includes SVM and ANN in terms of
Precision, Recall, and Dice Score Coefficient. Here 150
samples are taken as input for evaluating the performance and
result shows that the DCNN achieves higher true positives The
precision, recall and DSC give substantially more relevant
results for DCNN architecture.

TABLE 3. COMPARISON OF DCNN PERFORMANCE WITH EXISTING

DCNN Architecture Patch Size Accuracy (%) CLASSIFICATION ALGORITHMS

16×16 74.68
Classification Precision Recall Dice Score
DCNN Architecture 1 24×24 75.76 Algorithms (%) (%) Coefficient (%)

32×32 70.91
SVM 82.66 81.76 78.69
16×16 71.11
ANN 83.33 82.57 79.71
DCNN Architecture 2 24×24 73.25
32×32 73.16 DCNN(Architecture 1)
85.33 84.79 82.57
(24×24)
16×16 82.42
DCNN Architecture 3 DCNN(Architecture 2)
24×24 82.55 84.51 83.45 81.65
(24×24)
32×32 80.02

803
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE
DCNN(Architecture 3)
85.79 85.24 83.14
(24×24)
Table 3 results comparison of proposed DCNN and the
existing techniques includes SVM and ANN in terms of
Precision, Recall, and Dice Score
Figure 4. Comparison with Precision
*- DCNN Architecture with patch size 24×24.
Figure 4 shows the precision of different classification
algorithms, the proposed DCNN architectures are evaluated
using the input patch size of 24×24. From the figure that the
proposed architecture gives higher precision values than the
existing algorithm and the computes the values are close the
truth value
Figure 5. Comparison with Recall
*- DCNN Architecture with patch size 24×24
Figure 5 and Figure 6 shows the graphical representation
of recall and DSC values of different classification algorithms
804
of the figure that the performs better than the existing
algorithm.
Fig. 6. Comparison with Dice Similarity Coefficient
V. CONCLUSION AND FUTURE WORK
The detection and classification of lung tumor can greatly
increases the early stage detection of lung cancer and improves
patient survival. The designing and analyzing three different
architectures of DCNN for the lung tumor classification of
without computing texture features and the morphology. Three
architectures were defined with different number of layer end
up with the fully connected layer and the soft max classifier.
The original images are partition into smaller patches, and
used as the input to the first convolutional layer. The tumor as
well as non-tumor images are used to train the DCNN
architectures. For training and testing the LIDC-IDRI dataset
is best result of the architecture 3 with precision of 85.79 % ,
accuracy of 82.55%, recall of 85.24%, and dice score
coefficient of 83.14 % which outperforming the results
obtained by other architectures. In future the DCNN
architecture design can be applied for other parts of human
such as liver, brain and breast.
REFERENCES
[1] Gonzalez.R.C and Woods, R. E. (2009). Digital Image Processing.
(3nd ed). Addison-Wesley Longman Publishing Co., Inc.
[2] William D. Travis, MD, Elisabeth Brambilla, “The 2015 World Health
Organization Classification of Lung Tumors Impact of Genetic,
Clinical and Radiologic Advances”, Journal of Thoracic Oncology,
Volume 10, Number 9, September 2015.
[3] Devinder Kumar, Alexander Wong, David A. Clausi, “Lung Nodule
Classification Using Deep Features in CT Images”, 978-1-4799-1986-
4/15, 2015 IEEE, DOI 10.1109/CRV.2015.25.
[4] Ahamed, A. M. U., Eswaran, C., & Kannan, R. (2017). CBIR system
based on prediction errors. J. Inf. Sci. Eng, 33(2), 347-365.
[5] Emmanuel Adetiba and Oludayo O. Olugbara,” Lung Cancer
Prediction Using Neural Network Ensemble with Histogram of
oriented Gradient Genomic Features”, Hindawi Publishing
Corporation e Scientific World Journal Volume 2015, Article ID
786013, 17 pages.
[6] Khaing Zin Htwe,Kunihito YAMAMORI, Tin Mar Kyi, “Automated
Lung Nodule Classification by Artificial Neural Network and Fuzzy
 2019 International Conference on Computational Intelligence and Knowledge Economy (ICCIKE)
December 11–12, 2019, Amity University Dubai, UAE

Inference System” 978-1-5090-2333-2/16, 2016 IEEE 5th Global

Conference on Consumer Electronics.
[7] Jinsa Kuruvilla, K. Gunavathi, “Lung cancer classification using
neural networks for CT images”, Computer methods and programs in
bio medicine 113 ( 2

805