Lung Cancer Detection Using Wavelet Scattering Transform and Artificial Intelligence Technique

doi:10.21203/rs.3.rs-2494893/v1

Download PDF

Research Article

Lung Cancer Detection Using Wavelet Scattering Transform and Artificial Intelligence Technique

https://doi.org/10.21203/rs.3.rs-2494893/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The diagnosis of lung cancer is practically done by looking at a sample of lung cells in the lab. However, signs of lung cancer can be found by screening using X-Ray, CT, or Histopathological images. Each of these imaging modalities has its advantages and disadvantages. Chest X-ray is the first-line investigation for suspected lung cancer in primary care. However, the highest-quality studies suggest that the sensitivity of chest X-rays for symptomatic lung cancer is only 77–80%. On the other hand, a chest CT scan uses x-rays to make detailed cross-sectional images of the chest. Instead of taking 1 or 2 pictures, like a regular x-ray, a CT scanner takes many pictures and a computer then combines them to show a slice of the part of the chest under investigation. A CT scan is more likely to show lung tumors than traditional chest x-rays. In addition, the size, shape, and position of any lung tumors can be shown by a chest CT scan. More lung cancers were detected in the CT screening group compared with the control group with a 95% confidence interval. Moreover, Histopathological image analysis is widely used for cancer grading. Compared to mammography, CT and others, histopathology slides provide more comprehensive information for the diagnosis, and the diseases are analyzed by detecting tissue and cells in lesions. However, an invasive biopsy is necessary, which is often tried to be avoided. Therefore, chest CT is an optimal candidate for our study in sense of accuracy and availability. In this proposal, we deal with multi-class cancer detection from CT Lung images that is by detecting the cancer type rather than two classes (cancerous and normal images). The proposed method is based on a better representation of the image features by using Wavelet scattering Transform (WST). The classification is performed using three machine learning (ML) algorithms including support vector machine (SVM), kernel nearest neighbor (KNN), and random forest (RF). The WST coefficients are stable under signal deformations and globally invariant to signal translation and rotation. Based on the simulation results, the proposed method achieved an accuracy of 93.24%, 95.28%, and 99.90% for the case of WST + SVM, WST + KNN, and WST + RF networks, respectively.

Lung cancer Detection

Wavelet Scattering Transforms

Machine Learning

Chest-CT

Lung cancer is the leading cause of cancer-related deaths worldwide [1]. Early detection and precise diagnosis are crucial for successful treatment. The current diagnosis is made by looking at lung cells in the lab. The cells can be taken from lung secretions such as mucus, fluid removed from the area around the lung (thoracentesis), or from a suspicious area using a needle or surgery (biopsy). However, this method is sometimes invasive method.

On the other hand, image analysis techniques with different imaging modalities presented a comfortable way for cancer detection. Histological imaging of hematoxylin and eosin-stained tissue specimens remains the gold standard for lung cancer diagnosis [2]. However, tumor classification is challenging because of tumor heterogeneity. Also, chest X-rays, chest CT, MRI, and PET are used for lung cancer detection.

A CT scan [3] is more likely to show lung tumors than routine chest x-rays [4]. It can show the size, shape, and position of lung tumors. Low dose computed tomography (LDCT) is used as well in the detection of lung cancer. However, one of its drawbacks is the relatively high false positive rate [5].

Like CT scans, MRI scans show detailed images of soft tissues in the body. But MRI scans use radio waves and strong magnets instead of x-rays. MRI can detect and stage lung cancer, and this method could be an excellent alternative to CT or PET/CT in the investigation of lung malignancies and other diseases [6].

On the other hand, in a PET scan, a slightly radioactive form of sugar is injected into the blood and collects mainly in cancer cells to be traced. In addition, a combination of PET/CT using a special machine that can do both at the same time is used for better accuracy of cancer detection. This lets the doctor compare areas of higher radioactivity on the PET scan with a more detailed picture on the CT scan [7].

In sense of classifier/detector, there are many methods that appeared in the literature for lung cancer detection. These methods are mainly based on machine learning (ML) or deep learning networks (DLN) [8–11]. In [8], ML is used for cancer detection from histopathological images. Seven classifiers: naive Bayes, support vector machines (SVM) with Gaussian kernel, SVM with linear kernel, SVM with the polynomial kernel, bagging for classification trees, random forest utilizing conditional inference trees, and Breiman's random forest. The best accuracy obtained with this method is 85% by utilizing SVM with Gaussian kernel, random forest utilizing conditional inference trees, and Breiman's random forest.

In [9], a cancer diagnostic model based on a deep learning-enabled support vector machine (SVM) is proposed. This method used a convolutional neural network (CNN) with an SVM classifier. The authors reported an accuracy of 94% for pulmonary nodule detection representing early-stage lung cancer.

In [10], deep learning has been proposed as a promising tool to classify malignant nodules. The Lung Cancer Prediction Convolutional Neural Network (LCP-CNN) has been trained to generate a malignancy score for each nodule using CT data. The reported accuracy is 94.5%. Review in [11], provides a comprehensive review of algorithms and techniques relevant to major processes in the pipeline, including feature representation, feature indexing, searching, etc.

From these methods, it can be concluded that ML achieved better accuracy when combined with CNN and a feature extractor this is because ML is feature dependent. However, features extracted from CNN or hand-crafted features can be affected by image rotation zooming, and or scaling. Therefore, the accuracy is still low. Moreover, methods in the literature classify images into cancer or normal image without classifying to the type of lung cancer. Therefore, in this work, we propose a multi-class ML lung cancer detector that can classify the CT images into Adenocarcinoma, Large cell carcinoma, Squamous cell carcinoma, and Normal. The proposed method is based on Wavelet Scattering Transform, which is scale-independent, rotation independent and zoom independent. The proposed method is tested with different classifiers including SVM, Kernel Nearest Neighbor, and Random Forest. Our main contributions to this work are:

Unlike the previous work [13], we propose to classify CT lung images into four classes: Adenocarcinoma, Large cell carcinoma, Squamous cell carcinoma, and Normal.
Unlike the work done in [5], we used the wavelet scattering transform as a feature domain rather than the time domain to make the system more robust against scaling and shifting.

WST is used as a feature domain to extract features that are insensitive to scaling and shifting rather than using the time domain.

2.1 Preprocessing

The block diagram of the proposed system is shown in Fig. 1. The proposed system starts with the dataset partitioning into training, validation, and test dataset. For convenience, these images are transformed into 1D signals.

2.2. Wavelet scattering Transform

To compute the WST of a signal $x$, this signal is processed in three successive operations to generate wavelet scattering coefficients in each stage. These operations are convolution, nonlinearity, and averaging respectively. Figure 2 displays multiresolution/multilayer wavelet scattering transforms in which the scattering coefficients should be determined at each layer [12] as follows:

The zeroth-order scattering coefficients are calculated using basic input averaging as follows:

$${S}_{0}=x*\varphi , \left(2\right)$$

where $x$ is the input signal, $\varphi$ is the scaling function, and * is the convolution operator.

The high-frequency bands are captured by convolving with the mother wavelet ${{\psi }}_{{{\lambda }}_{\text{i}}}$at scale ${{\lambda }}_{\text{i}}$. So, the first-order scattering coefficients, S1, are generated by averaging the modules of the lowes band at the first scale in the filter bank ${{\lambda }}_{1}$.

$${S}_{1}x\left(t,{\lambda }_{1}\right)=\left|x\star {\psi }_{{\lambda }_{1}}\right|\star \varphi \left(3\right)$$

In the same way, the second wavelet transform is determined as follows:

$${S}_{2}x\left(t,{\lambda }_{1},{\lambda }_{2}\right)=\mid \mid x\star {\psi }_{{\lambda }_{1}}\left|\star {\psi }_{{\lambda }_{2}}\right|\star \varphi \left(4\right)$$

At each step, the signal at the lowest band incurs modules non-linearity and is averaged through the convolution by the father wavelet (low pass) $\varphi$ filter as shown in Fig. 1. For the m-th layer, the scattering coefficients, ${S}_{m}$, have been computed as follows.

$${S}_{m}x\left(t,{\lambda }_{1},\dots ,{\lambda }_{m}\right)=\mid \mid \mid x\star {\psi }_{{\lambda }_{1}}|\star \dots |\star {\psi }_{{\lambda }_{m}}\mid \star \varphi \left(5\right)$$

2.3 Artificial Intelligence

Machine learning is a branch of artificial intelligence (AI). The main aim of machine learning is to comprehend the multi-cancerous structure of data and fit it into models that people can comprehend and use. A solid prognosis requires fewer medical tests, which makes it important for ML techniques to perform well overall, offer medical professionals interpretable prognostic information, and enhance decision-making. The main drawback of the ML technique is the requirement for a robust feature extractor. Medical images in spatial domain are not sufficient as a feature for the ML techniques. The results of using the ML technique are not acceptable as in [13]. Therefore, the main clue in this proposal is to present WST as a feature domain to be combined with ML to classify CT lungs into multi-cancerous classes rather than cancerous and nun-cancerous images.

On the other hand, deep learning (DL) is a branch of machine learning. Despite DL doesn’t require feature extraction, it requires a huge number of datasets to be efficient. Therefore, DL has been combined with ML to take the benefit of the DL feature extractor and the ML classifier [5]. Even so, the performance of DL-ML classifiers is dependent on the architecture of the DL network. Transfer learning is used also to reduce the required dataset.

2.4 ML Classification Based on WST

As rotation-independent, scaling-independent, and shift-independent, WST is a good candidate for feature extraction. As has been mentioned, ML classification is highly dependent on the extracted features and its ability to represent the data and differentiate between various classes.

3.1. Dataset

Data contain three chest cancer types: Adenocarcinoma, Large cell carcinoma, and Squamous cell carcinoma, in addition to the normal cell. Lung adenocarcinoma is the most common form of lung cancer accounting for 30% of all cases overall and about 40% of all non-small cell lung cancer occurrences. Adenocarcinomas are found in several common cancers, including breast, prostate, and colorectal. Adenocarcinomas of the lung are found in the outer region of the lung in glands that secrete mucus and help us breathe. Symptoms include coughing, hoarseness, weight loss, and weakness [2].

Large-cell carcinoma lung cancer grows and spreads quickly and can be found anywhere in the lung. This type of lung cancer usually accounts for 10–15% of all cases of Non-Small Cell Lung Cancer. Large-cell carcinoma tends to grow and spread quickly.

Squamous cell carcinoma lung cancer is found centrally in the lung, where the larger bronchi join the trachea to the lung, or in one of the main airway branches. Squamous cell lung cancer is responsible for about 30% of all non-small cell lung cancers and is generally linked to smoking. The last type of images is the normal CT-Scan images. Three examples for each class are shown in Fig. 3.

The kernel function used with SVM is Gaussian. The number of classes is four. For the WST, the invariance scale is set to 32, and the sampling frequency is 128.

3.2. Implementation Environment

The system used for simulation is CPU Core I5, 2.6 GHz processor on a Windows 10 operating system. The software on which the model is implemented in MATLAB 2022a.

3.3. Performance Metrics

For performance measurement, multiple metrics are utilized, including confusion chart, accuracy, precision, recall, and F1-score. The confusion chart shows the success for each class and the confusion with other classes.

The accuracy of the model describes how well the model performs across classes. Accuracy is formulated as shown below. The accuracy can be calculated directly from the confusion chart by dividing the sum of the diagonal by the total number of samples

$$Accuracy= \frac{Number of trully classified samples}{Total number of samples}$$

Precision is the measure of the models’ capability to identify true positives, and it is calculated as shown below. The precession can be calculated from the confusion chart for each column by dividing the value at the diagonal in this column by the summation of that column.

$$Precision= \frac{True Positive}{True Positive+False Positive}$$

The recall is the ratio between the true positive prediction values and the sum of predicted true positive and false negative values. It is calculated as shown below. The recall can be calculated from the confusion chart for each row by dividing the value at the diagonal in this row by the summation of that row.

$$Recall= \frac{True Positive}{True Positive+False Negative}$$

F1 score is the overall model accuracy that balances precision and recall in a positive class. It is calculated as represented by:

$$F1 Score= 2\times \frac{Precession\times Recall}{Precession+Recall}$$

3.4 Experimental Results and Discussion

In this section, we present the simulation results for the CT lung cancer classification using the proposed system compared with the ML method [13], DL + ML method [5], and the DL networks including Darknet-19, VGG-19, and Alexnet. The comparison is based on four metrics: accuracy, precession, recall, and F1-score.

For comparison, the results of works done in [5, 13] are used for different metrics. For the DL networks, pre-trained networks (Darknet-19, VGG-19, and Alexnet) are used and the last three layers are retrained for learning transfer.

Table 1 shows the resulting evaluation metrics using different methods. From this table it can be shown that DL networks are efficient in detecting normal CT images with high precession (100% for the three tested networks), however, they achieved a low accuracy for the overall classification due to the confusion between different types of cancerous images (75.24%, 70.68% and 65.15% for Darknet-19, VGG-19, and Alexnet DL networks, respectively). Also, the combination of SVM as an ML for two states classification [13], results in better accuracy for the two states than using ML with special domain features.

On the other hand, using WST transform results in better image representation and differentiation between different cancerous images. Therefore, the combination of WST with the ML networks achieves high accuracy (93.24%, 95.28%, and 99.90% for the case of WST + SVM, WST + KNN, and WST + RF networks, respectively) with high precession for the four classes.

Another illustrative metric is the confusion chart that can explicitly show the number of confusions with different classes. The confusion charts for different methods are shown in Fig. 4 and Fig. 5. Figure 4 shows that for the three tested networks, the 92 normal images are classified correctly to be normal so that the precession for the normal CT image classification is 100%. However, there are 4 cancerous images confused with normal images to be detected as normal in the case of the Darknet-19 DL network. Also, in the case of the VGG-19 DL network, there are 2 cancerous images classified by mistake as normal. The same case is done with Alexnet Dl network, where 4 cancerous images are classified as normal due to the confusion between image features. Moreover, DL networks can lead to disastrous results due to feature confusion as in the case of Alexnet while classifying SCC images. The SCC features are confused with ACA features by using the Alexnet network which leads to 10% precession.

The high-performance achievement of the WST is due to multiscale contractions, linearization of hierarchical symmetries, and sparse representations. Moreover, WST coefficients are stable under signal deformations and globally invariant to signal translation and rotation.

In contrast, the proposed WST-based feature extraction achieved less confusion for all cancerous and normal images, as shown in Fig. 5. It is worth mentioning that the number of data is multiplied by the number of sub-bands.

Table 1

the evaluation metrics for different classification methods
Approach	Classifier	Class	Accuracy (%)	Precession (%)	Recall (%)	F1-Score (%)
Proposed	WST + SVM	ACA	93.24	95.43	89.37	92.30
		LCC		86.45	95.60	90.79
		Normal		98.29	99.21	98.75
		SCC		91.32	92.44	91.88
	WST + KNN	ACA	95.28	94.98	94.14	94.56
		LCC		92.11	96.38	94.20
		Normal		99.08	98.91	98.99
		SCC		94.99	93.22	94.09
	WST + RF	ACA	99.90	100	100	100
		LCC		99.92	99.55	99.73
		Normal		100	100	100
		SCC		99.68	99.94	99.81
Deep Learning Networks	DarkNet-19	ACA	75.24	70.83	75.22	72.96
		LCC		64.71	73.33	68.75
		Normal		100	95.83	97.87
		SCC		74.44	66.34	70.16
	VGG-19	ACA	70.68	78.33	75.20	76.73
		LCC		96.08	46.67	62.82
		Normal		100	97.87	98.92
		SCC		31.11	93.33	46.67
	AlexNet	ACA	65.15	94.17	55.12	69.54
		LCC		62.75	71.11	66.67
		Normal		100	95.83	97.87
		SCC		10	100	18.18
[5]	SVM + CNN		94	95	94.5	94.5
[13]	SVM		90	93	66	77

This work is applying various machine-learning techniques for lung cancer detection with cancer-type identification. The proposed method is based on the combination of independent transformation that is wavelet scattering transform with the ML classifiers to classify the chest-CT images into four classes. The four classes are Adenocarcinoma, Large cell carcinoma, Squamous cell carcinoma, and Normal. Three classifiers are tested with the WST: SVM, KNN, and RF. The proposed model WST + RF outperforms all other methodologies. The performance enhancement of the proposed model is due to the ability of the WST to understand the dataset and extract features from it without being affected by the imaging condition variation. Unlike the work done in the literature, the proposed model can differentiate between the different types of lung cancers effectively.

Ethical Approval

Authors declare that the manuscript consists of public dataset for cancerous and non-cancerous lung CT images and doesn’t include any human or animal studies.

Competing interests

Authors declare that the manuscript doesn’t include competing financial or personal interests.

Authors' contributions

Osama Omer, wrote the main manuscript text, implement the algorithm and prepared all the figures. Yoshifumi Saijo reviewed the manuscript.

Funding

Authors declare that they didn’t receive any fund for this work.

Barta J. A., Powell C. A., Wisnivesky J. P. Global epidemiology of lung cancer. Annals of Global Health. 2019;85(1):p. 8. doi: 10.5334/aogh.2419
Xu R, Wang Z, Liu Z, Han C, Yan L, Lin H, Xu Z, Feng Z, Liang C, Chen X, Pan X, Liu Z. Histopathological Tissue Segmentation of Lung Cancer with Bilinear CNN and Soft Attention. Biomed Res Int. 2022 Jul 7;2022:7966553. doi: 10.1155/2022/7966553. PMID: 35845926; PMCID: PMC9283032.
Tina D. Tailor, Edward F. Patz JrLung. Cancer Screening with Chest CT: Efficacy Confirmed. HomeRadiology: Imaging Cancer, Vol. 2, No. 3, May 2020, https://doi.org/10.1148/rycan.2020204015
Bradley SH, Abraham S, Callister ME, Grice A, Hamilton WT, Lopez RR, Shinkins B, Neal RD. Sensitivity of chest X-ray for detecting lung cancer in people presenting with symptoms: a systematic review. Br J Gen Pract. 2019 Nov 28;69(689):e827-e835. doi: 10.3399/bjgp19X706853. PMID: 31636130; PMCID: PMC6805164.
Snowsill T, Yang H, Griffin E, Long L, Varley-Campbell J, Coelho H, Robinson S, Hyde C. Low-dose computed tomography for lung cancer screening in high-risk populations: a systematic review and economic evaluation. Health Technol Assess. 2018 Nov;22(69):1–276. doi: 10.3310/hta22690. PMID: 30518460; PMCID: PMC6304730.
Hochhegger B, Marchiori E, Sedlaczek O, Irion K, Heussel CP, Ley S, Ley-Zaporozhan J, Soares Souza A Jr, Kauczor HU. MRI in lung cancer: a pictorial essay. Br J Radiol. 2011 Jul;84(1003):661–8. doi: 10.1259/bjr/24661484. PMID: 21697415; PMCID: PMC3473490.
Ohno Y, Koyama H, Nogami M, Takenaka D, Yoshikawa T, Yoshimura M, et al. STIR turbo SE MR imaging vs. coregistered FDG-PET/CT: quantitative and qualitative assessment of N-stage in non-small-cell lung cancer patients. J Magn Reson Imaging 2007;26:1071–80
Yu KH, Zhang C, Berry GJ, Altman RB, Ré C, Rubin DL, Snyder M. Predicting non-small cell lung cancer prognosis by fully automated microscopic pathology image features. Nat Commun. 2016 Aug 16;7:12474. doi: 10.1038/ncomms12474. PMID: 27527408; PMCID: PMC4990706.
Shafi, I.; Din, S.; Khan, A.; Díez, I.D.L.T.; Casanova, R.d.P.; Pifarre, K.T.; Ashraf, I. An Effective Method for Lung Cancer Diagnosis from CT Scan Using Deep Learning-Based Support Vector Network. Cancers 2022, 14, 5457. https://doi.org/10.3390/cancers14215457.
Marjolein A. Heuvelmans, Peter M.A. van Ooijen, Sarim Ather, Carlos Francisco Silva, Daiwei Han, Claus Peter Heussel, William Hickes, Hans-Ulrich Kauczor, Petr Novotny, Heiko Peschl, Mieneke Rook, Roman Rubtsov, Oyunbileg von Stackelberg, Maria T. Tsakok, Carlos Arteta, Jerome Declerck, Timor Kadir, Lyndsey Pickup, Fergus Gleeson, Matthijs Oudkerk. Lung cancer prediction by Deep Learning to identify benign lung nodules. Lung Cancer, Volume 154, 2021, Pages 1–4.
Zhongyu Li, Xiaofan Zhang, Henning Müller, Shaoting Zhang, Large-scale retrieval for medical image analytics: A comprehensive review, Medical Image Analysis, Vol. 43, 2018, pp. 66–84.
L. Sifre and S. Mallat, "Rotation, scaling and deformation invariant scattering for texture discrimination," in Proceedings of the IEEE conference on computer vision and pattern recognition, 2013, pp. 1233–1240.
Magdy, E.; Zayed, N.; Fakhr, M. Automatic classification of normal and cancer lung CT images using multiscale AM-FM features. Int. J. Biomed. Imaging 2015, 2015.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Lung Cancer Detection Using Wavelet Scattering Transform and Artificial Intelligence Technique

Status:

Version 1

Abstract

Figures

1. Introduction

2. The Proposed Multi-class Ml System

2.1 Preprocessing

2.2. Wavelet scattering Transform

2.3 Artificial Intelligence

2.4 ML Classification Based on WST

3. Results Of Ct Lung Cancer Detection

3.1. Dataset

3.2. Implementation Environment

3.3. Performance Metrics

3.4 Experimental Results and Discussion

4. Conclusion

Declarations

References

Additional Declarations

Status:

Version 1