Lung Cancer Detection System Using Deep Learning

Abstract—

Deep Learning methods are becoming very popular in medical applications due to their high reliability and ease. Lung Cancer is the type of cancer that is responsible for death of human being in the world. Lung cancer detection is performed by using various data analysis and classification techniques since the detection of cancerous nodule become impossible. This review considers approaches to build an automated diagnose system based on the lung cancer diagnosis task. 3D convolution and recurrent neural networks are used to detect cancerous nodule. Finally, We are comparing the overall accuracy of the neural networks on lung cancer detection problems.

Keyword – deep learning, convolution neural network, lung cancer detection system Index Terms—component, formatting, style, styling, insert

I. Introduction

Medical image processing with automatic diagnosis is one of the important tasks which can improve the healthcare field and save millions of patients life worldwide. It is the reason why there are a lot of efforts put into creation of the systems which can help doctors to improve the performance and accuracy of the medical diagnosis. In this work, we consider lung cancer detection problem. Lung cancer is a dangerous lung tumor featured by unconstrained cell growth in tissues of the lung. Lung cancer patients have very low life span in the advanced stages, where it is directly proportional with its growth at its detection time. There are multiple existing techniques that have been used for diagnosing lung cancer such as Computed Tomography (CT), Chest Radiography (x-ray), Magnetic Resonance Imaging (MRI scan) and Sputum Cytology.

This paper involves – sections. Section II and Section III gives a brief introduction to Artificial Intelligence techniques namely Artificial Neural Network and Convolution Neural Network. Section IV involves a detailed Literature Survey which presents the researches done by various authors in the area of Lung Cancer Detection using above mentioned techniques. Section V presents the problem definition for lung cancer detection system based on literature survey. In this paper Section VI defines AI techniques future scope.

II. Artificial Neural Network

Networks are nothing but the collection of multiple nodes that releases biological neurons of the human brain. Neurons Identify the applicable funding agency here. If none, delete this. are inter connected by links and they interact with each other. Nodes are used to take input and perform simple operations on data.

Perceptrons:

The perceptron is the most simple neural network which consists of a single neuron. Biological neurons have axons and dendrites. Like Biological Neurons, a perceptron is a single artificial neuron that consists of input nodes and a single output node which is connected to each input node.

Components of Artificial Neurons:

1) Input:

Each input node is associated with a numerical value, which can be any real number. Real numbers make up the full spectrum of numbers: they can be positive or negative, whole or decimal numbers.

2) Connections:

Similarly, each connection that departs from the input node has a weight associated with it, and this can also be any real number.

3) Next,

all the values of the input nodes and weights of the connections are brought together: they are used as inputs for a weighted sum: y=∑ D

i=1 wixi, or, stated differently, y=f(w1x1+w2x2+…wDxD). This result will be the input for a transfer or activation function.

4) Output:

The output node is associated with the function (such as the sigmoid function) of the weighted sum of the input nodes.

5) Bias:

Bias can be consider as the weight associated with an additional input node that is permanently set to 1. The bias value is critical because it allows you to shift the activation function to the left or right, which can make determine the success of your learning.

III. Main Importanat Data

We have planned our own model which relies on the reference paper we have read.

The overall architecture of the lung cancer detection system is shown above. This lung cancer detection system consists of the following steps:

1. Take a dataset of CT Scan Images.

The first step is to take a dataset that contains CT images of lungs. We are using Google Colab Environment for our project on which we have uploaded our Dataset. Our Dataset contains images in DICOM format. There are almost 200 images in our dataset and it consists of cancerous as well as non-cancerous images which will help us to easily diagnose Lung Cancer.

2. Image Acquisition

In this step, We retrieve the images from our dataset to perform further image pre-processing tasks. Image pre-processing is required to train, validate and test our System model.

3. Image Pre-processing

Digital image processing is nothing but the computer algorithms which perform image processing on digital images.

There are some image pre-processing techniques such as Noise Reduction, Binarization.

• Noise Reduction: Noise means, the pixels in the image show different intensity values instead of true pixel values that are obtained from the actual image. Noise reduction algorithm is the process of removing or reducing the noise from the image.
• Binarization: Binarization is nothing but converting a pixel image to a binary image.

4. Body region extraction

In this stage, we are performing segmentation on the images to find out the region of interest. In segmentation, we divide the entire image into small slices to check which region has the highest probability to get a malignant nodule. This task might be done by the segmentation algorithm RCNN(Recurrent Convolution Neural Network). Segmentation helps us in the detection of the malignant nodule. The accuracy of nodule detection highly depends on the working of the segmentation algorithm.

5. Feature Extraction

In this step, We extract the features from the obtained region which we get from the segmentation. The features are perimeter, area, entropy, density, irregularity index, contrast, correlation etc.

6. Classification

Classification play a major role in image processing techniques. It is used to classify the features that are taken from the images. They are classified into various classes based on their different properties. Deep learning uses different classification techniques. In this system, we are using CNN classification technique as CNN itself distinguish classes with a combination of algorithms.

IV. Literature Survey

Bohdan Chapaliuk, et.al. (2018) has developed a model which based on two approaches. First is 3D convolution networks which segment 3D patient scan and classification of the most mallignacy segments. The second one is to use a recurrent neural network to learn dependency between patient slices. There are two general approaches for distributed learning: model parallelism and data parallelism. There are different approaches to deal with neural networks parameters updates.

Fausto Milletari, et.al.has proposed a volumetric fully convolution network for the segmentation of 3D images. They have used MRI volumes of the prostate to tain the CNN and predict segmentation for the whole volume at once. They have introduced a special objective function that is optimised during training which is based on Dice Coefficient.They mainly deal with situations where there is a strong imbalance between foreground and background voxels. CNN is trained end-to-end on MRI volumes depicting prostate, and learns to predict segmentation for the whole volume at once. Wentao Zhu, et.al.(2018)has proposed a fully automated CT cancer dignosis system based on deep learning called DeepLung. DeepLung consists of two parts, nodule detection and classification. For nodule detection, they have designed 3D Faster R-CNN and U-net. Finally, gradient boosting machines with combined features are trained to classify nodules into malignant or non-malignant. The nodule classification subnetwork was validated on a public dataset from LIDC- IDRI, on which it achieved better performance than state-of-the-art approaches and surpassed the performance of experienced doctors based on image modality.

Fangzhou Liao, et.al. have evaluated the Malignancy of Pulmonary Nodules. The model consists of 2 modules. The first one is the 3D region proposal network for nodule detection which gives all unsure nodules as output. The second one selects the top five nodules based on detection confidence and evaluates their cancer probabilities. The overfitting caused by the shortage of the training data is alleviated by training the two modules alternately. The proposed model won first place in the Data Science Bowl 2017 competition. Rahul Dey, et.al. has developed a model which consists of two sequences GRU1 and GRU2 generated from the MNIST dataset and the IMDB dataset. The main driving signal of the gates appears to be the state as it contains essential information about other signals. It evaluates three variant models on MNIST and IMDB datasets of the Gated Recurrent Unit (GRU) in recurrent neural networks (RNNs) by retaining the structure and systematically reducing parameters in the update and reset gates.

V. Problem Definition

Lung cancer is the main reason of cancer death in the world. The experienced doctors take around 10 to 15 min to perform a detailed check for each patient because some nodules are small and hard to be found. Doctors can also evaluate the malignancy of nodules based on their structure, but the accuracy highly depends on doctors experience, and different doctors may give different predictions. Computer-Aided Diagnosis is suitable for this task because computer vision models can quickly scan everywhere with equal quality and they are not affected by fatigue and emotions. The recent advancement of deep learning has enabled computer vision models to help doctors to diagnose various problems.

VI. Conclusion

There are two approaches that are used to detect lung cancer: 1.3D Convolution neural network which is used for image processing i.e is segment 3D patient scan and classification of malignancy segments. 2. Use of the recurrent neural network to learn dependency between patient slices. We can acheive human level accuracy on the different medical images by using these approaches. Result shows that neural network which is trained to detect lung cancer on whole lung 3D image gives worse accuracy in comparision to two stages approach when two different neural networks for segmentation and classification are trained. Recurrent neural networks show competitive accuracy and performance.