Lung cancer recognition and prediction according to random forest ensemble and RUSBoost algorithm using LIDC data

Abstract

Lung cancer is increasing day by day. According to the report published by WHO in 2017, number of death people due to lung cancer in Bangladesh reached 12,075 or 1.53% of total deaths. It is a matter of great sorrow that this lung cancer detection does not occur at an early stage for many people. It is very important to recognize lung cancer and detect the location as well as an accurate prediction. Many researchers applied techniques like Fast Fourier Transform (FFT) for image enhancement, thresholding for the purpose of segmentation and binarization system for extraction, etc. Basically recognition process consists of three stages like enhancement, segmentation and feature Extraction, lung cancer images can be used as inputs and maintaining these stages give more quality and accuracy in the detection of lung cancer. Approaches developed by the earlier researchers fail to produce accuracy in real-time applications. Hence, in order to mitigate the drawbacks of these approaches a hybrid method to detect lung cancer. Gabor filter has been used for image enhancement of input image. Marker-Controlled Watershed algorithm has been applied for segmentation purpose which will help to provide the exact location of the infected region in the input pictures as well as Out-of-bag (OOB) error rate for different iteration (trees) has been observed graphically according to Random Forest Ensemble and OOB rate for the different class also been determined. When the number of tree is increasing, OOB rate in percentage is decreasing. The highest OOB value is 23.40% which provides an initial tree. Distribution curve of Mean decrease in accuracy, Mean decrease in Gini index and standard error of importance measure according to Random Forest Ensemble has also been observed graphically. RUSBoost algorithm has been introduced finally in order to evaluate accuracy. Performance of RUSBoost algorithm has been visualized in this paper for the purpose of Subtlety, Spiculation, Sphericity, Texture, Margin, Malignancy, Lobulation, where Texture provides the highest accuracy and Spiculation provides second highest accuracy.

Keywords

Gabor filter random forest RUSBoost OOB error rate

1. Introduction

The Lung analysis, short-term explanation of lung cancer, as well as its categories, are offered in this segment. The lungs are cone-shaped organs and a pair of sponge-like. The right lung is bigger than the left lung. The left lung has two lobes and the right lung has three lobes. Oxygen is carried into the lungs while air is breathing. Lung tissue conveyances oxygen to the bloodstream to drive to the relaxation of the body [1]. Cells relief Carbon dioxide by the way of the custom oxygen. The bloodstream transports carbon dioxide spinal to the lungs as well as the carbon dioxide shrubberies the body while air is breathing out. Lung cancer is an infection of irregular cells increasing and rising into a tumor. Cancer cells can be approved away from the lungs in lymph fluid, or blood that settings lung tissue. Metastasis happens as soon as a cancer cell foliage the place where it created and transfers into an alternative portion of the body or a lymph node via the bloodstream [2, 3]. Pair of lungs are the vital organs, which is the function of the respiratory system in the human body [4]. Continuing cigarette smoking is due to 85% majority of cases of lung cancer [5]. Approximately 10–15% never smoking who cases lung cancer. Non-small cell lung cancers (NSCLC) and small cell lung cancers (SCLC) are the two types of lung cancer which grow and types are different [6]. Lung cancer is the greatest deadly category of cancer and effective behavior depends on initial recognition [7]. The National Cancer Institute forecasts 228,190 original circumstances of lung beside bronchus cancer as well as 159,480 passing away, outstanding equally prostate cancers’ and breast expected deaths by means of further than 100,000 [8].

To struggle with this critical disease, radiologists practice CAD schemes by way of a second attitude to promotion with discovering and illustrating hypothetically cancerous nodules. In related work, a CAD scheme has been engaged as a second attitude for mutual residents, as well as the board, specialized radiologists. The CAD schemes measured in this paper help with the classification of lung nodes by calculating their semantic features. These semantic features consist of ratings for the classifications of Subtlety, Spiculation, Sphericity, Texture, Margin, Malignancy, Lobulation. To create these estimates, former research castoff region outlines physically calculated by radiologists by means of a foundation to arise the features low-level image [9]. After these features of an image, estimates of semantic features were prepared. To calculate if there is an essential for radiologists to stay with attraction physical separations for these estimates, we propose a CAD scheme constructed on several computer-derived weak segmentations (WSCAD) as well as illustration that its prognostic accuracy is as good as with the estimates achieved expending a CAD scheme established on the physical radiologists’ segmentation. Our offered scheme would additional subjective estimations with an exactly-defined algorithmic method. It also has significant cost-savings perspective, by replacing expensive time of radiologists with the petty cost of numerical calculation [10].

At first, this paper illustrated lung cancer recognition using Gabor filter according to the method of image processing than input several images and detected these lung is normal or abnormal (lung cancer). Section 3 described our proposed image processing method and the section of experimental result visualized lung cancer recognition using Gabor filter for image enhancement. Sections 4 and 5 described a theoretical description of Random forest ensemble and RUSBoost algorithm respectively. Experimental results visualized out-of-bag (OOB) error rate for different iteration (trees) according to Random Forest Ensemble and tabulation for Subtlety, Spiculation, Sphericity, Texture, Margin, Malignancy, Lobulation according to RUSBoost algorithm. At last visualized the accuracy of RUSBoost test and training for the purpose of Subtlety, Spiculation, Sphericity, Texture, Margin, Malignancy, Lobulation.

2. Literature review

Some researchers have offered and realized the detection of lung cancer using dissimilar methodologies of machine learning and image processing. Aggarwal et al. [10] offered a model that affords classification among the structure of normal lung anatomy and nodules. The scheme excerpts gray level, geometrical, and statistical features. LDA is castoff as classifier then optimum thresholding for subdivision. The scheme has 53.33% specificity 97.14% sensitivity and 84% accuracy. While the scheme identifies the nodule of cancer, its accuracy is quietly unacceptable. They have not been used any machine learning methods. They have been classified besides simple segmentation methods is used. Consequently, the arrangement of one of its stages in our novel model does not offer the probability of development. Jin et al. [11] used CNN as a classifier in his CAD scheme to identify lung cancer. The scheme has 86.7% of specificity, 82.5% of sensitivity and 84.6% of accuracy. The benefit of this ideal is that it customs spherical filter in Region of interest (ROI) taking out of phase which diminishes the training cost as well as detection steps. While application cost is cheap, it has quiet unacceptable accuracy. Sangamithraa and Govindaraju [12] proposed the unsupervised learning of $K$ mean system for segmentation before clustering. It clusters the pixel dataset along with definite features.

For cataloging these model apparatuses network for backpropagation. Structures like homogeneity, correlation, SSIM, entropy, PSNR, pull out via the gray-level co-occurrence matrix (GLCM) scheme. The scheme has around accuracy of 90.7%. Image preprocessing average filter is castoff for noise elimination which can be suitable for our novel model to eradicate the noise besides progress the accuracy. Roy et al. [13] established a scheme to notice nodule of lung cancer using fuzzy interfering scheme and dynamic contour exemplary. This scheme uses gray revolution for enhancement image dissimilarity. Image binarization is skilled earlier segmentation then ensued image is segmented via a dynamic contour model. Classification of Cancer is implemented using fuzzy implication scheme. Structures like entropy, correlation, mean, area, minor axis length, major axis length pull out to train all classifier. Ignatious and Joseph [14] established a scheme via watershed segmentation. In preprocessing it customs Gabor filter to improve the quality of the image. It associates the neural fuzzy model and its accuracy as well as region growing scheme. Gonzalez and Ponomaryvo [15] proposed a scheme that lung cancer classifies as malignant or benign. The scheme customs the prior info besides Housefield Unit(HU) to compute the Region of Interest(ROI). Shape structures corresponding area, fractal dimension circularity, eccentricity and textural structures similar to energy, mean, skewness, variance, contrast, entropy and softness are pull out to train and organize the SVM to recognize whether the lump is malignant or benign. Awai and his team [16] determined that the custom of a CAD scheme by way of a second attitude expressively helped the capability of both the residents and board specialized radiologists to identify chest pulmonary nodes in computed tomography (CT) scans.

3. Lung cancer recognition

Figure 1 illustrates lung cancer recognition system using Gabor filter for image enhancement.

Figure 1.

Lung cancer recognition using Gabor filter for image enhancement.

3.1 Image enhancement

Image enhancement methods can be categorized into two categories: spatial domain as well as Frequency domain. It supports to develop the opinion and interpretability of frontier areas in the image for humanoid observers. Varying of pixel rate benefits to system modifications in orthogonal distorted image otherwise, it offers enhanced processing procedures founded on frequency domain scheme it implements. On the other hand, preprocessing implements are recycled as image enhancement methods for additional image processing they are most suitable. FFT enhancement besides Gabor filtering are the three approaches castoff as image enhancement methods [17].

3.2 Image segmentation

Consuming this procedure maximum of the image exploring task can be prepared consequently. In exact, the prevailing approaches rest on extremely on the segmentation consequence for image explanation and detection. On the other hand, now we are consuming thresholding and Watershed segmentation. Achieved image subsequently segmentation from thresholding had considerably implication approximating fast treating speed fewer storage space than basically by operation of 512 grey levels image. Thresholding is the maximum leading implement for image segmentation by changing innovative pixel values by dark pixel rates. Thresholding opt for a threshold rate T besides it gives two stages to the image that is below value and beyond value for unusual threshold rate [18].

3.3 Gabor filter

Gabor filter is mainly used for the detection of edge that is a linear filter. In the longitudinal domain, 2D Gabor filter is called as Gaussian filter role modified by a function of sinusoidal [19]. In the procedure of this detection of cancer images, castoff is a 2D image, therefore expending 2D Gabor filter

$\displaystyle G(\sigma,\theta,\lambda,\psi,\gamma;a,b)=\exp\frac{-(a^{\prime 2% }+\gamma^{2}b^{\prime 2})}{2\sigma^{2}}\cos\left(2\frac{a^{\prime}}{\lambda}+% \psi\right)$ (1)

3.4 Region growing

Region growing is a method to prolong the enclosed area by assembling pixels otherwise sub-regions established on prearranged conditions. Essentially, the methodology is to begin a set opinion and at that time growing the area adjacent the kernel which has the similar properties by way of the kernel, for illustration, the region that has the similar range of gray level otherwise color [20]. Figures 2–6 display three steps of segmentation via region growing scheme.

3.5 Marker-controlled watershed

There are two key methods in segmentation, containing regional method and edge method. Watershed segmentation scheme relating mutually of these methods. The watershed scheme is an authoritative scheme to develop fast segmentation consequences. The basic indication of separation watershed alteration derives from topography, wherever an image is observed as CT image with gray level by means of an extended surface elevation [21].

4. Random forest

Random forest is also called random decision forests which are an ensemble learning scheme for regression, cataloging and supplementary tasks that control by creating an assembly of decision trees at training period then outputting the period that is the mode of mean prediction otherwise the classes of the different trees. Random decision forests are accurate for decision trees’ custom of their training set overfitting.

Random forest is a malleable, stress-free to a custom machine learning system that produces, uniform deprived of hyper-parameter alteration, an excessive consequence maximum of the period. It is moreover one of the greatest recycled processes, as it’s effortlessness and the element that it can be castoff for mutually regression tasks as well as classification [22, 23].

The training procedure for random forests relates the common method of bagging otherwise bootstrap aggregating to tree learners. Known a training set $A=a_{1},\ldots,a_{n}$ with reactions $B=b_{1},\ldots,b_{n}$ , bagging constantly ( $t$ times) chooses a random model with a standby of the training set then turns trees to these models:

For $t=1,\ldots,T$ :

1.
Model, with standby, $n$ training instances from $A$ , $B$ ; demand these $A_{t}$ , $Y_{t}$ .
2.
Train a regression or classification tree $f_{t}$ on $A_{t}$ , $B_{t}$ .

Algorithm:

Known: Set the examples $S$ is ( $a_{1},b_{1}$ ), …, ( $a_{n},b_{n}$ ) with minority class $b_{r}\in B$ , $|B|=$ 2

Weak Learner, WeakLearn

Iteration numbers, $T$

Preferred total instances’ percentage to be denoted by the class of minority, $N$

1.
Prepare $D_{1}(i)=\frac{1}{n}$ for every $i$ . 2.
Do on the behalf of $t=1,2,3,\ldots,T$

i
Build short-term training dataset $S^{\prime}_{t}$ using distribution $D^{\prime}_{t}$ via random undersampling
ii
WeakLearner is called, given that it with samples $S^{\prime}_{t}$ as well as $D^{\prime}_{t}$ is their weights
iii
Develop a hypothesis $h_{t}$ : $A\times B\rightarrow[0,1]$
iv
Determine the pseudo-loss (for $D_{t}$ and $S$ ):

$\displaystyle\!\!\!\!\!\!\!\!\!\!\varepsilon_{t}\!=\!\!\!\!\!\!\sum_{(i,b):y_{% i}\neq y}\!\!\!\!\!D_{t}(i)(1\!-\!h_{t}(a_{1},b_{1})\!+\!h_{t}(a_{1},b))$
v
Determine the weight inform parameter:

$\displaystyle\alpha_{t}=\frac{\varepsilon_{t}}{1-\varepsilon_{t}}.$
vi
Inform $D_{t}$

$\displaystyle\!\!\!\!\!\!\!\!\!\!\!\!D_{t+1}(i)\!=\!D_{t}(i)\alpha_{t}^{\frac{% 1}{2}(1+h_{t}(a_{i},b_{i})-h_{t}(a_{i},b:b\neq b_{i}))}$
vii
Stabilize $D_{t+1}$ : Let $Z_{t}=\sum_{i}D_{t+1}(i)$ .

$\displaystyle D_{t+1}(i)=\frac{D_{t+1}(i)}{Z_{t}}$

3.
The output of the ultimate hypothesis

$\displaystyle H(a)=\mathop{\arg\max}_{b\in B}\sum_{t=1}^{T}h_{t}(a,b)\log\frac% {1}{\alpha_{t}}$

Figure 2.
Lung cancer recognition using Gabor filter for image enhancement (result: detected normal lung).

Figure 3.
Lung cancer recognition using Gabor filter for image enhancement (result: detected lung cancer).

Subsequently training, estimates for hidden models $a^{\prime}$ can be prepared by being around the prediction or calculations from all the separate regression trees on $a^{\prime}$ :

$\displaystyle\hat{f}=\frac{1}{T}\sum_{t=1}^{T}f_{y}(a^{\prime})$ (2)

Otherwise, by enchanting the majority choose in the instance of an arrangement of trees.

Moreover, an approximation of the improbability of the calculation can be prepared as the standard deviation of the calculations after all the separate trees for regression on $a^{\prime}$ .

$\displaystyle\sigma=\sqrt{\frac{\sum_{t=1}^{T}(f_{y}(a^{\prime})-\hat{f})^{2}}% {T-1}}$ (3)

The quantity of examples/trees, $T$ , is a permitted parameter. Naturally, a little hundred to numerous thousand trees is recycled, reliant on the dimension of the training set. An optimum number of trees $T$ can be established consuming cross-validation, or by detecting the OOB error: the mean calculation or prediction error on individually training model $a_{i}$ , consuming merely the trees that did not have $a_{i}$ in their bootstrap model.
5. RUSBoost

Algorithm:
Known: Set the examples $S$ is ( $a_{1},b_{1}$ ), …, ( $a_{n},b_{n}$ ) with minority class $b_{r}\in B$ , $\|B\|=$ 2
Weak Learner, WeakLearn
Iteration numbers, $T$
Preferred total instances’ percentage to be denoted by the class of minority, $N$
1. Prepare $D_{1}(i)=\frac{1}{n}$ for every $i$ . 2. Do on the behalf of $t=1,2,3,\ldots,T$ i Build short-term training dataset $S^{\prime}_{t}$ using distribution $D^{\prime}_{t}$ via random undersampling ii WeakLearner is called, given that it with samples $S^{\prime}_{t}$ as well as $D^{\prime}_{t}$ is their weights iii Develop a hypothesis $h_{t}$ : $A\times B\rightarrow[0,1]$ iv Determine the pseudo-loss (for $D_{t}$ and $S$ ): $\displaystyle\!\!\!\!\!\!\!\!\!\!\varepsilon_{t}\!=\!\!\!\!\!\!\sum_{(i,b):y_{% i}\neq y}\!\!\!\!\!D_{t}(i)(1\!-\!h_{t}(a_{1},b_{1})\!+\!h_{t}(a_{1},b))$ v Determine the weight inform parameter: $\displaystyle\alpha_{t}=\frac{\varepsilon_{t}}{1-\varepsilon_{t}}.$ vi Inform $D_{t}$ $\displaystyle\!\!\!\!\!\!\!\!\!\!\!\!D_{t+1}(i)\!=\!D_{t}(i)\alpha_{t}^{\frac{% 1}{2}(1+h_{t}(a_{i},b_{i})-h_{t}(a_{i},b:b\neq b_{i}))}$ vii Stabilize $D_{t+1}$ : Let $Z_{t}=\sum_{i}D_{t+1}(i)$ . $\displaystyle D_{t+1}(i)=\frac{D_{t+1}(i)}{Z_{t}}$ 3. The output of the ultimate hypothesis $\displaystyle H(a)=\mathop{\arg\max}_{b\in B}\sum_{t=1}^{T}h_{t}(a,b)\log\frac% {1}{\alpha_{t}}$

Data sampling methods effort to improve the difficulty of class inequality by modifying the distribution of a class of the training data set. This can be skilled by either get rid of instances from the class of majority otherwise accumulation instances to the class of minority (oversampling). Some of the maximum mutual data sampling method is known as RUS. Contrasting additional sampling of complex data algorithms, RUS creates no effort to “intelligently” eradicate instances from the training information. As a substitute, RUS humbly eradicates instances from the class of majority at arbitrary till a distribution of preferred class is realized [24]. In this paper, mainly used RUSBoost algorithm for the purpose of prediction of lung cancer and determined the percentage of Margin, Spiculation, Sphericity, Lobulation, Lobulation, Malignancy, Texture for different countable instances’.

Figure 4.

Lung cancer recognition using Gabor filter for image enhancement (result: detected lung cancer).

Figure 5.

Lung cancer recognition using Gabor filter for image enhancement (result: detected lung cancer).

Figure 6.

Lung cancer recognition using Gabor filter for image enhancement (result: detected normal lung).

Figure 7.

Distribution of mean decrease in accuracy and mean decrease in Gini index according to random forest ensemble.

Figure 8.

Out-of-bag (OOB) error rate for different iteration(trees) according to random forest ensemble.

Table 1

Determine OOB for the purpose of Class 1 and Class 2 according to the number of trees

Number of trees	OOB	Class 1	Class 2	Number of trees	OOB	Class 1	Class 2
1	23.40%	28.57%	17.78%	16	12.40%	14.17%	12.77%
2	21.71%	25.33%	18.18%	17	12.80%	13.33%	12.31%
3	21.47%	22.83%	20.20%	18	12.00%	13.33%	10.77%
4	18.60%	19.23%	18.02%	19	12.00%	13.33%	10.77%
5	17.75%	16.36%	19.01%	20	10.80%	13.33%	8.46%
6	18.07%	18.42%	17.74%	21	11.20%	12.50%	10.00%
7	17.43%	15.65%	19.05%	22	10.40%	10.83%	10.00%
8	14.69%	17.09%	12.50%	23	10.40%	11.67%	9. 23%
9	14.17%	15.25%	13.18%	24	8.80%	9.17%	8.46%
10	13.36%	13.56%	13.18%	25	9.20%	10.83%	7.69%
11	12.05%	11.76%	12.31%	26	9.20%	11.67%	6.92%
12	14.80%	15.00%	14.62%	27	9.20%	10.00%	8.46%
13	12.00%	13.33%	10.77%	28	8.00%	10.83%	5.38%
14	13.20%	15.00%	11.54%	29	8.00%	9.17%	6.92%
15	11.60%	12.50%	10.77%	30	8.00%	10.00%	6.15%

6. Experimental results

6.1 Experimental results of lung cancer recognition

The procedure to acquire the separation via region growing technique is such as the subsequent Figs 2–6; initially, choice the region that will be the mark CT image, which is the left lung beside the right lung, now, set the kernel in this region. Moreover, expand the kernel size so as to it protections all the chosen regions. Figures 2–6 display the consequence of separation via region growing technique.

Figure 9.

Distribution of mean decrease in accuracy, mean decrease in Gini index and standard error of importance measure according to random forest ensemble (for test data).

To develop the segmentation consequence of watershed technique, the stages are as surveys; initially, determine the distance gradient on behalf of the detection of an edge, and formerly scratch the mark CT image by consuming a morphological method named initial via modernization. The exploration of three segmentation technique constructed on their achievement to novelty the segmentation consequences of the mark object of CT image. Figures 2–6 exposed two scheme have effectively created the mark object of CT images containing (a) region growing as well as (c) marker-controlled watershed through covering. Temporarily (b) marker measured watershed segmentation has failed to develop segmentation consequen-ces, subsequently, there are related adjacent to the mark object of CT image. Binarization is the method of varying the pixel values color into two periods, for example, black and white. Subsequently receiving the number of black and white pixels on separation effects, formerly we matched it with a value of threshold to calculate the situation of a lung (cancer or normal). The value of the threshold is acquired from explanations on a normal lung. The value of the threshold that is recycled in this exploration is 17179.

Figure 10.

Out-of-bag (OOB) error rate for different iteration(trees) according to random forest ensemble (for test data).

6.2 Experimental results of random forest ensemble

Figure 7 illustrated the distribution of Mean decrease in accuracy and Mean decrease in Gini index according to Random Forest Ensemble a Fig. 8 illustrated out-of-bag (OOB) error rate for different iteration(trees) according to Random Forest Ensemble than determined the OOB value for the purpose of several numbers of trees. Table 1 described that the number of tree is increasing, OOB value is decreasing. The highest OOB value is 23.40% which provides an initial tree.

Table 2
Tabulation for: Subtlety

Value	Count of instances	Percent
1	112	4.25%
2	244	9.26%
3	558	21.17%
4	903	34.26%
5	819	31.07%

Table 3

Tabulation for: Sphericity

Value	Count of instances	Percent
1	2	0.08%
2	143	5.42%
3	703	26.67%
4	1208	45.83%
5	580	22.00%

Table 4

Tabulation for: Margin

Value	Count of instances	Percent
1	110	4.17%
2	223	8.46%
3	414	15.71%
4	848	32.17%
5	1041	39.49%

Table 5

Tabulation for: Lobulation

Value	Count of instances	Percent
1	1409	53.45%
2	781	29.63%
3	313	11.87%
4	97	3.68%
5	36	1.37%

Table 6

Tabulation for: Spiculation

Value	Count of instances	Percent
1	1634	61.99%
2	637	24.17%
3	199	7.55%
4	116	4.40%
5	50	1.90%

Figures 8 and 9 visualized distribution plot of Mean decrease in accuracy, Mean decrease in Gini index and standard error of importance measure and Out-of-bag (OOB) error rate for different iteration(trees) according to Random Forest Ensemble along with test data set of LIDC. Where in Fig. 10 shows the number of tree is zero, the OOB error rate is 7000 then decreased OOB value for increasing number of tree (iteration).

6.3 Experimental results of RUSBoost

Figure 13 visualizes testing accuracy of RUSBoost algorithm for the purpose of Subtlety, Spiculation, Sphericity, Texture, Margin, Malignancy, Lobulation. Spiculation and Texture provides the highest accuracy on the other hand Sphericity provides an accuracy which is zero.

Table 7
Tabulation for: Texture

Value	Count of instances	Percent
1	171	6.49%
2	94	3.57%
3	156	5.92%
4	338	12.82%
5	1877	71.21%

Table 8

Tabulation for: Malignancy

Value	Count of instances	Percent
1	290	11.00%
2	581	22.04%
3	1211	45.94%
4	394	14.95%
5	160	6.07%

Figure 11.

Distribution of mean decrease in accuracy, mean decrease in Gini index and standard error of importance measure according to random forest ensemble (for training data).

Figure 12.

Number of tree vs. test classification error curve according to RUSBoost test.

Figure 13.

Testing accuracy of RUSBoost algorithm for the purpose of Subtlety, Spiculation, Sphericity, Texture, Margin, Malignancy, Lobulation.

Figure 14.

Training accuracy of RUSBoost algorithm for the purpose of Subtlety, Spiculation, Sphericity, Texture, Margin, Malignancy, Lobulation.

Figure 14 visualizes training accuracy of RUSBoost algorithm for the purpose of Subtlety, Spiculation, Sphericity, Texture, Margin, Malignancy, Lobulation. Where Texture provides highest accuracy and Spiculation provides the second highest accuracy. On the other hand, Sphericity also provides an accuracy which is zero.

7. Conclusion

In this paper, three image segmentation approaches have been implemented and evaluated for examining lung cancer, for example, Marker-Controlled Watershed, Region Growing, as well as Marker-Controlled Watershed with Covering. The consequences illustrate that Marker-Controlled Watershed with Covering provides the greatest enactment in term of separation consequence and organization time. Consequently, we choice Marker-Controlled Watershed with Covering scheme in image segmentation period. Moreover, in the extraction of the feature in a stage, we practice color feature for the exploration of lung cancer via binarization. Lastly, the binarization technique was an effectively resolute situation of the lung (cancer or normal) from the CT scan image. This paper also implemented Random forest ensemble and RUSBoost using test and train LIDC data set. Lobulation. Spiculation and Texture provide the highest accuracy for testing dataset according to RUSBoost. On the other hand, Texture provides the highest accuracy and Spiculation provides the second highest accuracy for training dataset.

References

Aung

H.H.

Sivakumar

Gholami

S.K.

Venkateswaran

S.P.

Gorain

and Md

, An overview of the anatomy and physiology of the lung, in: Nanotechnology-Based Targeted Drug Delivery Systems for Lung Cancer, 2019, pp. 1–20. https://doi.org/10.1016/B978-0-12-815720-6.00001-0.

Moulla

Gradistanac

Wittekind

Eichfeld

Gockel

and Dietrich

, Predictive risk factors for lymph node metastasis in patients with resected non-small cell lung cancer: A case control study, Journal of Cardiothoracic Surgery (2019). https://doi.org/10.1186/s13019-019-0831-0.

Schuchert

M.J.

et al., Factors influencing recurrence following anatomic lung resection for clinical stage I non-small cell lung cancer, Lung Cancer 128 (2019), 145–151. https://doiorg/10.1016/j.lungcan.2018.12.026.

Bharati

Podder

Mondal

Mahmood

and Raihan-Al-Masud

, Comparative Performance analysis of different classification algorithm for the purpose of prediction of lung cancer, in: Intelligent Systems Design and Applications, 2018, pp. 1–11. https://doi.org/10.1007/978-3-030-16660-1_44.

Chan

E.D.

, Vulnerability to nontuberculous mycobacterial lung disease or systemic infection due to genetic/heritable disorders, in: Nontuberculous Mycobacterial Disease. Respiratory Medicine Griffith

, ed., Humana Press, Cham, 2019. https://doi.org/10.1007/978-3-319-93473-0_4.

Gemine

R.E.

et al., Longitudinal study to assess impact of smoking at diagnosis and quitting on 1-year survival for people with non-small cell lung cancer, Lung Cancer 129 (2019), 1–7. https://doiorg/10.1016/j.lungcan.2018.12.028.

et al., A neuro-heuristic approach for recognition of lung diseases from X-ray images, Expert Systems with Applications 126 (2019), 218–232. https://doiorg/10.1016/j.eswa.2019.01.060.

Ries

et al., SEER Cancer Statistics Review, 1975–2005, National Cancer Institute, Bethesda, MD, 2008, pp. 1975–2005.

Loyman

, and Greenspan

, Lung nodule retrieval using semantic similarity estimates, in: Proc SPIE 10950, Medical Imaging 2019: Computer-Aided Diagnosis, 109503P 2019. doi: 10.1117/12.2512115.

10.

Aggarwal

Furqan

and Kalra

, Feature extraction and LDA based classification of lung nodules in chest CTscan images, in: 2015 International Conference On Advances in Computing, Communications and Informatics (ICACCI), 2015, DOI: 10.1109/ICACCI.2015.7275773.

11.

Jin

Zhang

and Jin

, Pulmonary nodule detection based on CT images using convolution neural network, in: 2016, 9Th International Symposium On Computational Intelligence and Design (ISCID), 2016, doi: 10.1109/ISCID.2016.1053.

12.

Sangamithraa

and Govindaraju

, Lung tumour detection and classification using EK-Mean clustering, in: 2016 International Conference On Wireless Communications, Signal Processing and Networking (Wispnet), 2016, doi: 10.1109/WiSPNET.2016.7566533.

13.

Roy

Sirohi

and Patle

, Classification of lung image and nodule detection using fuzzy inference system, in: International Conference On Computing, Communication and Automation, 2015, doi: 10.1109/CCAA.2015.7148560.

14.

Ignatious

and Joseph

, Computer aided lung cancer detection system, in: 2015 Global Conference On Communication Technologies (GCCT), 2015, doi: 10.1109/GCCT.2015.7342723.

15.

Rendon-Gonzalez

and Ponomaryov

, Automatic lung nodule segmentation and classification in CT images based on SVM, in: 2016 9Th International Kharkiv Symposium On Physics and Engineering of Microwaves, Millimeter and Submillimeter Waves (MSMW), 2016, doi: 10.1109/MSMW.2016.7537995.

16.

Awai

Murao

Ozawa

Komi

Hayakawa

Hori

and Nishimura

, Pulmonary nodules at chest CT: Effect of computer aided diagnosis on radiologists’ detection performance, Radiology 230(2) (2004).

17.

Ayshath Thabsheera

A.P.

Thasleema

T.M.

and Rajesh

, Lung cancer detection using CT scan images: A review on various image processing techniques, in: Data Analytics and Learning. Lecture Notes in Networks and Systems. vol. 43 Nagabhushan

Guru

Shekar

and Kumar

, eds, Springer, Singapore, 2019, pp. 413–419, https://doi.org/10.1007/978-3-030-05348-2_1.

18.

Sehad

Chibani

Hedjam

et al., Gabor filter-based texture for ancient degraded document image binarization, Pattern Analysis and Applications 22(1) (2019), 1–22. https://doiorg/10.1007/s10044-018-0747-7.

19.

Kavitha

M.S.

Shanthini

and Sabitha

, ECM-CSD: An efficient classification model for cancer stage diagnosis in CT lung images using FCM and SVM techniques, Journal of Medical Systems (2019). https://doi.org/10.1007/s10916-019-1190-z.

20.

Bhatia

Sinha

and Goel

, Lung cancer detection: A deep learning approach, Soft Computing for Problem Solving (2019), 699–705. https://doi.org/10.1007/978-981-13-1595-4_55.

21.

Braun

et al., Single chromosome aneuploidy induces genome-wide perturbation of nuclear organization and gene expression, Neoplasia 21(4) (2019), 401–412. https://doi.org/10.1016/j.neo.2019.02.003.

22.

Paul

et al., Improved random forest for classification, IEEE Transactions on Image Processing 27(8) (2018), 4012–4024.

23.

Bharati

Rahman

M.A.

and Podder

, Breast cancer prediction applying different classification algorithm with comparative analysis using WEKA, in: 2018 4th International Conference on Electrical Engineering and Information and Communication Technology (iCEEiCT), Dhaka, Bangladesh, 2018, pp. 581–584. doi: 10.1109/CEEICT.2018.8628084.

24.

Seiffert

et al., RUSBoost: A hybrid approach to alleviating class imbalance, IEEE Transactions on Systems, Man, and Cybernetics 40(1) (2010), 185–197.