Professional Documents
Culture Documents
Abstract
Lung cancer is one type of cancer that first appears in the lungs. Cancer arises when the body's cells
begin to multiply uncontrollably. In order to discover lung cancer and localize the tumour, a variety
of procedures, including CT scans, MRIs, and chest x-rays, were carried out. To save as many lives
as possible, early cancer identification is now crucial. Because there is so much noise in medical
imaging, it can be difficult to predict what will happen next. The main objective of this study is to
develop an automated system for detecting lung cancer using the transfer learning. a deep learning
method on the VGG16 architecture, while ensuring that no crucial information is lost during image
processing on the data collected from significant, qualified centers with professional training.
1. Introduction
When cells produce too many new cells and the older or aberrant ones don't die when they have
to, then cancer develops. Cancer cells may displace healthy cells due to their uncontrolled
development. Due to this, our body struggles to operate normally. Cancer comes in a variety of
forms. Cancer is called by the area of the body where it first appeared and can appear anywhere
in the body. A tumor is a growth or mass. However, the majority of bumps are not cancerous.
Benign refers to lumps that are not cancerous. Malignant lumps are cancerous growths. Cancer
differs from benign tumors in that it can spread to other regions of the body. Benign tumors
cannot. Cancerous cells are capable of separating from the cancer's primary growth location.
These cells may spread throughout the body and risk ending up in the lymph nodes or in other
organs, which might interfere with their ability to carry out their usual duties [1].
Lung cancer is one form of cancer which begins in lungs is lung cancer. When the body's cells
start to proliferate out of control, cancer develops [2]. Numerous tests, including CT scans,
MRIs, and chest x-rays, were performed to find lung cancer so that doctors could locate the
tumor and treat it appropriately. Today, early cancer detection is essential to saving as many
lives as possible. It's challenging to anticipate what will happen next because of the significant
noise present in medical imaging. As a consequence, medical pictures will be preprocessed using
image processing techniques, and then lung cancer will be detected using machine learning and
deep learning algorithms based on preprocessed image segmentation.
The major goal of this study is to create an automated system for detecting lung cancer using
deep learning technique called transfer learning on VGG16 architecture, while making sure that
no essential information is lost during image processing on the data acquired from significant,
professionally educated authorized centers.
2. Literature survey
Our survey indicates that there are many various types of imagery utilized for medical imaging,
but computer tomography (CT) scans are typically chosen because of the lower noise. The best
technique for feature extraction, object categorization, and medical imaging has been
demonstrated to be deep learning. Related work to categorize lung cancer is shown in table 1.
Instead than using a set of pre-programmed instructions, deep learning algorithms will
automatically analyze patterns from massive volumes of data [3]. A Convolutional Neural
Network (ConvNet/CNN) [5] is the Deep Learning method which takes in the input picture, gives
various elements and objects in the image priority (learnable weights and biases), and distinguish
between them. Comparatively speaking, the ConvNet requires substantially less pre-processing
than other classification techniques. ConvNets has the capacity to learn these properties and
filters, when it comes to basic techniques filters were hand-engineered.
Structure of the ConvNet is similar to connection pattern of neurons in human brain and was
modelled after how visual cortex is organized. Only in Receptive Field, the constrained area of
visual field, do individual neurons react to inputs. Such fields are grouped together and
encompass the full visual field. A sample CNN architecture is shown in figure 1.
Convolutional layer, pooling layer, ReLU correction layer, and fully-connected layer are the 4
different types of layers that make up a convolutional neural network.
3. 1 Convolutional layer:
The convolutional layer, that is always first layer, serves as the brain of convolutional neural
networks. Its objective is to discover the presence of a particular set of features in the input
images. To achieve this, convolution filtering is employed. The fundamental concept is to "drag"
the window which represent the feature across the image, calculate the convolution product
between feature and each segment of the scanned image, and then move the window away from
the image.
Multiple images are therefore supplied into the convolutional layer, which then calculates the
convolution for every image with each filter. The features in the images match the filters
precisely. Then receive the feature map to each image and filter combination which shows where
features were located in image. The greater the number, the more the associated location in the
image matches the feature. The network learns features during the training phase rather than
having them defined beforehand. Weights of the convolution layer are referred to as filter kernels.
Gradient descent back propagation is used to initialize them and subsequently update them [6].
3.2 Pooling layer
This sort of layer, which receives many feature maps and executes the pooling operation for each
of them, is frequently sandwiched between two layers of convolution. The pooling procedure
involves reducing size of photos while maintaining their crucial features. The amount of feature
maps we receive as output and input are equal, but they are substantially smaller. Network's
parameters and computations are simplified by pooling layer. This increases the network's
effectiveness and prevents over learning [7].
The ReLU correction layer substitutes 0s for the negative values which are provided as inputs. It
carries out the task of activation [8,9]. The real non-linear function is known as ReLU (Rectified
Linear Units) is defined as follows:
ReLU(x)=max(0,x) (1)
Here,
x : input value
Larger value between 0 and the input value is the ReLu activation function's output value.
The last convolutional or pooling layer's output is sent to fully connected layer, and it is flattened
prior to being used. Last layer, which follows the fully connected layers, employs the softmax
activation function [10] to figure out probability of which input belongs to a certain class
(classification).
4. Methodology
CNN-based image classification has demonstrated reasonably excellent performance. The goal
of this research study is to employ Transfer learning [11], a deep learning approach, to classify
lung cancer images.
Lung cancer dataset [12] from “Iraq-Oncology Teaching Hospital/National Center for Cancer
Diseases (IQ-OTH/NCCD)”, which have patients CT images with different stage lung cancer
and also healthy volunteers. Oncologists and radiologists at these 2 centers marked
“IQ-OTH/NCCD” slides. 1190 pictures reflect CT scan slices from 110 patients in the collection.
Malignant, benign and normal cases are divided into three categories. There are 40 cases that
have been identified as malignant, 15 that have been identified as benign, and 55 that have been
classed as normal. Four case studies, including malignant and non-malignant pictures. CT scans
are acquired in DICOM format at first. Scanner utilized is a Siemens SOMATOM.
The CT procedure involves the following steps:
Sample images of normal and malignant cases is shown in Figure 2 and Figure 3.
With base dataset, base network was trained, and then features learnt from first stage were reused
for the other network to train on the second dataset and task. The use of pre-trained models on
comparable data has yielded positive results in image classification tasks. Few firms had
produced models that require weeks to train on contemporary technology, such as the famous
CNN’s are Microsoft ResNet Model [15], Google Inception Model [14], and Oxford VGG
Model [13]. Those models may be downloaded, used with newer models that use images as input
to provide more accurate result.
4.3 Oxford VGG Model
At “Image Net Large Scale Visual Recognition Challenge (ILSVRC)” in 2014, VGG shown in
Figure 4, won 2nd place in the image classification and first place in image localization test. This
model was built by an Oxford team of researchers, who made weights and structure available
online. Structure was created by just 3*3 convolution layers, 2*2 max pool layers, and fully-
connected layers at the last, as shown in Fig. 4. The input picture should be 224*224*3 pixels in
size (RGB image).
The images in the collection were all resized to 512x512 pixels, and the dataset includes normal
and malignant cases at various phases of development. There were 763 images in train set, 152
images in validation set, and 188 images in test set.
Image will be loaded into the VGG16 layer, which will include pre train weights, and fully
connected dense layer with sigmoid activation as final layer. Suggested system, as illustrated in
Figure 5, is made up of 2 layers: pre-trained VGG16 layer and the dense layer. Weights that have
been pre-trained for VGG16 model will be imported, summary of the model before training is
shown in figure 6.
Figure 6. Summary of the model
Input data will be learned with pre-trained weights, and single learning layer will be dense layer
with the sigmoid activation function and does a good job with binary classification. For
compilation, to calculate loss, binary_crossentropy and ‘adam’ optimizier were used. And model
is trained till 30 epochs, and batch size as 32.
1
s(x) = −x (2)
1+ e
Here,
e : Euler’s number
x : Dot product of each neuron value with weights
N M
1
Logloss = -
N
∑ ∑ y ij log (¿ ¿ pij )¿ ¿ (3)
i j
Here,
Input Dataset
Data Collection
Normal
Malignant
Data pre-processing
Proposed model
VGG16
Dense layer
with sigmoid
Compile
Binary cross
entropy
Adam
optimizer
Training Classification
Figure 5: Proposed deep-Architecture to classify lung cancer
5. Experimental results
The experiment was carried out in google collab, and used a free GPU to train the model.The
measurements recorded during the training of dataset were loss and accuracy. These
measurements were measured for both training and validation data. VGG16 model has achieved
100 percent training accuracy, where as for validation accuracy, VGG16 achieved 84 percent
accuracy. VGG16 correctly detects lung cancer at early, medium and final stage and also it
detects normal cases. In the next work, needed to improve the validation accuracy by working
with different deep learning models, and select which suits the best deep learning model to detect
lung cancer early. Loss graph and accuracy graph for the proposed model were shown in figure
7. The qualitative results predicted by the proposed system is displayed in table 2.
Figure 7: Loss and Accuracy graph
For early lung cancer detection, proposed an deep architecture based on transfer learning,
modified VGG16 model with dense layer as sigmoid activation function, and achieved 100
percent training accuracy, 83 percent as validation accuracy. VGG16 correctly detects lung
cancer at early, medium and final stage and also it detects normal cases. In the next work, needed
to improve the validation accuracy by working with different deep learning models, and select
which suits the best deep learning model to detect lung cancer early. We will train different deep
learning architectures like VGG19, ResNet50 and Inception V3, and compare the results among
those architecture and identify which suits the best for lung cancer detection.
References
[1] https://www.cancer.org/treatment/understanding-your-diagnosis/what-is-cancer.html
[2] https://www.cancer.org/cancer/lung-cancer/about/what-is.html
[3] Schmidhuber, Jürgen. "Deep learning in neural networks: An overview." Neural networks 61
(2015): 85-117.
[4] Phung, Van Hiep, and Eun Joo Rhee. "A high-accuracy model average ensemble of
convolutional neural networks for classification of cloud image patches on small datasets." Applied
Sciences 9.21 (2019): 4500.
[5] O'Shea, Keiron, and Ryan Nash. "An introduction to convolutional neural networks." arXiv
preprint arXiv:1511.08458 (2015).
[6] Albawi, Saad, Tareq Abed Mohammed, and Saad Al-Zawi. "Understanding of a convolutional
neural network." 2017 international conference on engineering and technology (ICET). Ieee, 2017.
[7] Sun, Manli, et al. "Learning pooling for convolutional neural network." Neurocomputing 224
(2017): 96-104.
[8] Ramachandran, Prajit, Barret Zoph, and Quoc V. Le. "Searching for activation functions." arXiv
preprint arXiv:1710.05941 (2017).
[9] Agarap, Abien Fred. "Deep learning using rectified linear units (relu)." arXiv preprint
arXiv:1803.08375 (2018).
[10] Sharma, Sagar, Simone Sharma, and Anidhya Athaiya. "Activation functions in neural
networks." towards data science 6.12 (2017): 310-316.
[11] Torrey, Lisa, and Jude Shavlik. "Transfer learning." Handbook of research on machine learning
applications and trends: algorithms, methods, and techniques. IGI global, 2010. 242-264.
[12] https://www.kaggle.com/datasets/hamdallak/the-iqothnccd-lung-cancer-dataset
[13] Simonyan, Karen, and Andrew Zisserman. "Very deep convolutional networks for large-scale
image recognition." arXiv preprint arXiv:1409.1556 (2014).
[14] Szegedy, Christian, et al. "Rethinking the inception architecture for computer
vision." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.
[15] He, Kaiming, et al. "Deep residual learning for image recognition." Proceedings of the IEEE
conference on computer vision and pattern recognition. 2016.
[16] Ausawalaithong, Worawate, et al. "Automatic lung cancer prediction from chest X-ray images
using the deep learning approach." 2018 11th Biomedical Engineering International Conference
(BMEICON). IEEE, 2018.
[17] Haarburger, Christoph, et al. "Image-based survival prediction for lung cancer patients using
CNNS." 2019 IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019). IEEE,
2019.
[18] Sori, Worku J., et al. "DFD-Net: lung cancer detection from denoised CT scan image using
deep learning." Frontiers of Computer Science 15.2 (2021): 1-13.
[19] Alam, Janee, Sabrina Alam, and Alamgir Hossan. "Multi-stage lung cancer detection and
prediction using multi-class svm classifie." 2018 International Conference on Computer,
Communication, Chemical, Material and Electronic Engineering (IC4ME2). IEEE, 2018.
[20] Kakeda, Shingo, et al. "Improved detection of lung nodules on chest radiographs using a
commercial computer-aided diagnosis system." American Journal of Roentgenology 182.2 (2004):
505-510.
[21] Cheran, Sorin Cristian, and Gianfranco Gargano. "Computer aided diagnosis for lung CT using
artificial life models." Seventh International Symposium on Symbolic and Numeric Algorithms for
Scientific Computing (SYNASC'05). IEEE, 2005.
[22] Causey, Jason L., et al. "Lung cancer screening with low-dose CT scans using a deep learning
approach." arXiv preprint arXiv:1906.00240 (2019).