Detection and classification of liver cancers using computed tomography images

doi:10.21203/rs.3.rs-2242020/v1

Download PDF

Research Article

Detection and classification of liver cancers using computed tomography images

https://doi.org/10.21203/rs.3.rs-2242020/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Objectives

The aim of this research was to develop an appropriate algorithm-routine that can automatically detect and classify liver cancers from the digital images of CT sample based on the texture and morphological features.

Method

The digital image processing (DIP) provides techniques first for liver image segmentation using semi-automatic technique from the enhanced abdominal CT images and classification using Artificial Neural Network (ANN) through texture features, then the liver tumor automatically extracted from the liver image using Otsu method which is a histogram based thresholding and the classification using ANN through morphological features.

Result

The samples were collected from a number of thirty six diagnosed patients in the procedure of abdominal CT images, at axial view in Black Lion Hospital from DICOM store in Addis Ababa. These images were selected to be processed and analyzed to get best generated result of the predefined classification model based on the extracted features. First nine texture features differentiate 12 normal and 24 abnormal liver images out of the given sample’s containing 36 images then tumor images extracted from the 24 abnormal images, the ANN classifier using six morphological features classified the 24 abnormal liver images as benign and malignant. It’s result presented by confusion matrix and showed that the maximum accuracy rate for tumor classification was 95.8%. The performance can be increased more by increasing the number of samples.

Conclusions

The outcome of this research work help radiologists to follow the condition of diseases at early stages and classify tumor as benign and malignant. It also helps in decreasing the rate of presence of disease by detecting cancer in its earlier stages.

Nuclear Medicine & Medical Imaging

Digital image processing (DIP)

Liver cancer and Computed Tomography

Liver cancer that is caused by uncontrolled growth of liver cells might result in a lump called tumor.Tumor is a pathologic disturbance of cell growth, characterized by excessive and abnormal proliferation of cells which may be solid or fluid filled [1].

Liver cancer may originate from the liver itself (called primary) or may come from some other organs of the body (called secondary). There are various types of tumors that may be found inside liver: benign, malignant, and metastases.

A benign tumor is non-cancerous, not posing serious threat to health and are quite common and does not spread to other parts of the body. The most common types of benign tumors are hemangioma and cysts. whereas a malignant tumor also called primary liver cancer, is cancerous, originates in the liver, grows inside the liver or on the surface and can be spread to the other part of organs in the body. It presents symptomatically as an increase in abdominal mass, abdominal pain, jaundice, nausea or liver dysfunction. The most common examples of malignant tumors are cholangio-carcinoma and hepatocellular carcinoma (HCC). The chronic liver disease, HCC (hepatoma) is caused either by virus or alcohol infection. Such infections on the liver referred as hepatitis, with the most common of which are hepatitis A, B, and C. Hepatitis accounts for 75% of all liver cancers [2].

Liver cancer that begins in another organ and spread to the liver as a new site, usually through the blood stream called the secondary cancer or Metastases. Some common examples of liver metastases are colorectal cancers, adenocarcinoma from pancreas and stomach, bronchial cancer, endocrine and breast cancers. Surgical resection is the only curative treatment for metastases [3].

Malignant tumors such as HCC, have a poor prognosis, with few treatments offering long term survival. If it is not detected at an early stage, HCC has indeed a poor prognosis with only 5%, 5 year survival [4, 5].

The only treatments with a curative value are surgical resection, liver transplantation and percutaneous ablation. Surgical resection is indeed available for less than 10% of patients. Moreover, survival after the resection is low for HCC with 5 year survival rate that ranges 20–40%; depending on the histology and the extension of the lesions [5].

According to Ethiopian Journal of Health Development [6], Ethiopia is as far behind in addressing hepatocellular-carcinoma (HCC), which is chronic liver disease with 100% fatality rate as compared to other countries. Patients with HCC cases die within few months following diagnosis. In Ethiopia there is a lack of information on the burden risk factors, diagnosis modalities, surveillance strategies and treatment approaches to HCC.

One of the most important aspects of liver diseases prevention is the classification management of potential risk factors. Public health efforts to reduce the toll of these diseases has to be considerably hampered if information on treatable risk factors is unavailable. These issues create the need of a computer aided or diagnostic system that utilizes a better, accurate and effective classification systems.

Various medical imaging modalities can be employed for the diagnosis of liver cancers such as US, CT, MRI, and angiography. However CT preferred over those modalities because of it’s high accuracy, sensibility and specificity, it is non-invasive, quick usually take 10–20 min, may be done with or without contrast, produce detailed images, painless and generally safe, provide high resolution image with an accurate detection [7]. Also CT images can give specific information about the size, shape, and location of any tumors in the liver. It can be used to guide a biopsy needle precisely into a suspected tumor [8]. Hence, this work investigated the possibility of identification of tumor in the liver using CT images and also classify one of the prevalent liver disease, primary liver cancer effectively & accurately.

This research work was designed to develop computer vision techniques for detection and classification of liver cancer by using image processing and neural network toolboxes in Matlab software in which the proposed methodology shown below in Fig. 1.

Image Acquisition

In this section only abdominal computed tomography (CT) images were focused on for image acquisition The data’s was collected from Black Lion Hospital in DICOM store in Addis Ababa. One slice in axial view per diagnosed patient from 128 slices of images that clearly shows liver was taken by the help of Radiologist.

Before segmentation process, erosion was used for data compression to remove unwanted background, using the “diamond” structure element with a ‘radius of 2’ as proposed by Sahoo et al., [9].Then filtering operation using a median filter was also performed to reduce noise present in it. The median filter used a neighborhood (kernel) of 5×5 pixels around the evaluated pixel and this value was obtained from the previous work of Malik et al., [10].

Liver Segmentation

the liver is both of highly variable shape and it displays low difference appearance compared to the neighboring organs, segmentation of the liver is a very difficult task because of two reasons; First, the liver is a soft organ, whose shape is highly dependent on the other organs in the abdomen. Moreover, many pathologies have a strong effect on the appearance and the shape of the liver. For instance, no edges are visible on many sides of the liver. In particular the differences in between the liver and the diaphragm or the stomach or the spleen are very small [11].

For this work, first the liver image was segmented from the enhanced abdominal CT image by using semi-automatic technique which involves in separating the unwanted part by a correct cut-off non-liver region without losing pixels that belong to the liver.

Texture Feature Extraction

After segmentation of liver from abdominal CT image, texture features were designed to extract nine texture features. Texture analysis was comprising by two stages: image filtration, followed by quantification of texture. It is more efficient to employ the filter in the frequency domain, as convolution of the filter mask and the image in the spatial domain is equivalent to multiplication of the Fourier transforms of the filter mask and the image in the frequency domain.Texture features were quantified as mean, variance, standard deviation, range, contrast, energy, homogeneity, correlation and entropy.

To obtain textural features, co-occurrence matrix adopted first. The co-occurrence matrix method of texture description based on the repeated occurrence of gray level configuration in the texture. An occurrence of a gray level configuration described by a matrix of relative frequencies P_θ,d (i,j), giving how frequently two pixels with gray levels appear in the window separated by a distance d in direction θ and these matrices are symmetric, p_θ,d = P_θ,d (i,j).The GLCMs are normalized by dividing each entry of the matrices by the total number of pairs as follows:

p(i,j) = P (i,j) /n (3.1)

Where, p(i,j) is the normalized GLCM and

P (i,j) - is the un-normalized GLCM and n is the normalization factor, which is equal to the total number of pairs. Texture features measured using GLCM derived quantities by taking the ratio of each normalized GLCM elements multiplied by a logarithmic function of fine and medium texture respectively. These values could be considered to represent the relative contributions made by each fine and medium [12]. This texture ratio quantification was used because the ratio would effectively normalized, thereby minimizing the effects of any potential variations in CT attenuation values occurring from one image to another and also reducing the effect of noise on texture quantification.

The procedure adopted in obtaining these features given in equations (3.3) through (3.11).

Where; p(i,j) is gray level co-occurrence matrix, µ - mean, σ - standard deviations, var -variance, Con - contrast, EG - energy, HOM - homogeneity, CO - correlation, EN - entropy &

TR-texture ratio. For all the separated gray level components, the derivation of co-occurrence matrices, P_θ,_d(i,j) were done for different directions. For all pixels (i₁,j₁) in the image, (i₂,j₂) at a distance, d in direction, θ were determined.

Classification of Liver

Since we had two classes that correspond to the normal and abnormal, we used two binary bit numbers. If the value of the column vector is 0, it indicates that the feature data set is the member of the normal liver class and if it is 1, it indicates that the feature data set is the member of abnormal liver class.

In this research work, classification of liver was designed to differentiate normal liver and abnormal liver images by the texture of liver images using a neural network classifier.

A classifier always requires a training and testing data to train and evaluate performance of the classifier. The sample data’s for training liver images obtained from radiologist who has assigned 24 images from abnormal and 12 images from normal, a total of 36 liver images.

Nine texture features namely mean, standard deviation, variance, range, contrast, energy, homogeneity, correlation, and entropy were selected as input to the network. Accordingly, in this study, the input layer has nine nodes and the hidden layer has ten nodes and the output layer has two nodes.

Liver Tumor Segmentation

Otsu thresholding which is histogram based, was used for tumor segmentation due to its simplicity and the difference in the gray levels intensities, which requires different threshold for different images. It works based on pixel distributions and similarity among the pixels within a region.

The histogram shown in Fig. 4, is suitable for image thresholding.The right bottom region of image was discarded as this region normally not contain the tumor. A histogram of the image is analyzed and the highest pitch represents the middle intensity of tumor region.

In this section, segmentation of liver tumor performed by Otsu method which is histogram based thresholding. The valley point is usually chosen as the threshold. In bi-level thresholding, all the gray level values greater than the threshold values were assigned the tumor label and all the other gray level values were assigned the background label. Thus the tumor pixels were separated from the background pixels.

Where T is the threshold. Where T is the threshold. If the same value of T is adopted for the entire image, the process is called global thresholding. When the choice of value for T at coordinates point (x,y) depends on statistical properties of pixel values in a neighborhood around (x,y) referred as regional thresholding. Then after thresholding method, area opening technique was used to remove the connected components(objects) from a binary image that have pixels lower than a set value based on the procedures suggested by Al Mahmud et al., [13]. Then, pixel subtraction operator was used by taking two images as input and producing as output a third image.The pixel values of the output were simply the pixel values of the first minus the pixel values of the second image [14].

R(i, j) = S(i, j) – T(i, j) (3.13)

Where R, S and, T represent the output image, thresholded image and the image after small object removal, respectively.

Morphological Feature Extraction

The classification system proposed are based on the morphology of the image by shape. The image characteristics of the liver tumor can be obtained from the analysis of binary images. The following morphological features were extracted from the binary images to classify liver tumor images: area, perimeter, major and minor axes lengths, eccentricity and roundness of the tumor.

Liver Tumor Classification

Since we had two classes that correspond to malignant & benign and, we used two binary bit numbers. If the value of the column vector is 0, it indicates that the feature data set is the member of the benign class and if it is 1, it indicates that the feature data set is the member of malignant. A classifier always requires a training and testing data to evaluate classifier performance. The sample data’s for training obtained from the 24 segmented images of abnormal liver tumors: 12 images from malignant and 12 images from that of benign were assigned by radiologist. The six morphological features were selected as an input to the Neural network. Accordingly, in this study, the input layer has six nodes and the hidden layer has ten nodes and the output layer has two nodes. The neural network trained by adjusting the weights so as to be able to predict the correct class. Classification accuracy was observed in MSE and % E. If the results of MSE and % E are small in the samples given for training, testing and validation, the value is observed in either confusion matrix analysis or ROC. Only confusion matrix method was considered in this work.

Analysis of the Classifier

In order to evaluate the effectiveness of the proposed algorithm to classify the liver tumors, four metrics were used.These are: true positives (TP), true negatives (TN), false positives (FP), false negatives (FN). TP represent mass lesion marked as malignant, which is also classified as malignant. TN represent mass lesion marked as benign, which is also classified as benign. FP mass lesion marked as malignant, but classified as benign. FN represents the mass lesion marked as benign, but classified as malignant.

Table 1

Confusion matrix
	Positive	Negative
Positive	TP	FP
Negative	FN	TN

The mathematical equations that were used to evaluate the performance of the classifier for tumor classification described as follows.

Sensitivity = TP /TP + FN (3.14)

Specificity = TN /TN + FP (3.15)

Accuracy = TP + TN /TP + TN + FP + FN (3.16)

In this section, the results of detection, extraction and classification of liver and tumor are presented.

Pre-Processing Results

According to the medical ethics, for keeping low profile of patients, the original abdominal CT image which is not shown in this research was undergoes an Erosion techniques by an operator having ‘diamond’ structure element with a ‘radius’of 2 to compress non-interesting information of the CT image such as (Right/Left, Posterior/Anterior, Name, Examination date, Label and CT marks shown in Fig. 7a). Then the image undergoes a median filtering to increase it’s enhancement which is shown in Fig. 7b. It was performed with a median operator of kernel of 5×5 pixels size by going through each element of pixel of the image and replaces with the median of its neighboring pixels which were located in a neighborhood (kernel) of 5×5 pixels around the evaluated pixel.

Liver Segmentation Result

The steps for segmentation result of normal and abnormal liver images shown in Fig. 8a and b, respectively. It was performed by semi-automatic techniques which creates a binary mask from the ROI (liver) with all pixel values inside the liver to 1 and all pixel values outside the ROI to 0. Finally, the liver was delineated from other part of the CT image by green component segmentation.

Texture Based Liver Classification Result

The result of texture feature extraction of liver images shown below in Table 2.

Table 2

Texture features sample
Texture features
Liver images	Mean	Variance	Standard deviation	Range	Contrast	Energy	Homogeneity	Correlation	Entropy
Normal	1020	1.1869e + 07	3.4452e + 03	27679	0.0335	0.2539	0.9832	0.9935	1.298338
Abnormal	1029	1.2704e + 07	4.0148e + 03	23044	0.1611	0.1911	0.6209	0.6769	1.183054

Tumor Segmentation Result

After applying histogram based thresholding using Otsu method as mentioned in section 3.4, to the filtered gray scale image, the threshold value turned grayscale image into a binary image and noticeable that the tumor edge was correctly separated from other parts of the liver image. This technique allows the detection of the tumor as a separate object in the image. The result of liver tumor segmentation using different techniques with different steps tested through different types of images are shown below in Fig. 10a & b.

Morphological Extraction Result

The result of Morphological feature extraction of liver tumor images shown below in Table 3.

Table 3

Morphological features sample extracted
Morphological features
Liver tumors	Area	Perimeter	Major axis	Minor axis	Eccentricity	Roundness
Malignant	1320	63.44	17.48	12.04	0.47	0.50
Benign	716	76.32	15.47	9.31	0.75	0.85

Tumor Classification Result

The result of Confusion matrix using morphological features shown below in Table 4.

Table 4: Confusion matrix using morphological features

Performance of Classifier

Table 5

Performance result of classifier
	TP	TN	FP	Sensitivity (%)	Specificity (%)	Accuracy (%)
Training	10	6	0	100	100	100
Validation	1	2	1	100	66.7	75
Test	1	3	0	100	100	100
All	12	11	1	100	91.7	95.8

In this section, the discussions of detection, extraction and classification of liver and tumor are presented.

As shown from Table 2, the abnormal liver image has lower value in energy, homogeneity, correlation,

and entropy and higher value in contrast than that of the normal liver images. This indicate that there are some variation in neighboring pixels value in a gray level image of abnormal liver images compared to that of the normal one as a result, the texture of abnormal liver image became an irregular and a blurred appearance. First the liver image was enhanced by removing the noise using median filtering and converted to binary image using thresholding technique then area opening technique was used to clean the image by eliminating the unwanted components that have lower pixel values than a set value. Finally the tumor was segmented by subtracting cleaned image from the binarized one by using the Image subtraction operator. From the selected sample of malignant and benign liver tumors shown in Table 3, benign liver tumor has circular shape or more rounded and eccentric geometry than malignant liver tumor.

As shown in Table 4, the summary result of ANN classifier; using morphological features: for the total training of 16 samples (10 malignant & 6 benign), all instances were correctly classified, and for the total validation of 4 samples (2 malignant & 2 benign), 3 images were correctly classified and 1 image misclassified, for the total test of 4 samples (1 malignant & 3 benign), all instances were correctly classified, and from all total samples of 24 images (12 malignant & 12 benign) 23 images (95.8%) were correctly classified and 1 image (4.2) was misclassified.

From the performance result of the classifier shown in Table 5, the sensitivity, specificity and the overall accuracy ot the classifier were: 100%, 100%,100% respectively for training; 100%, 66.7%,75% respectively for validation, 100%,100%,100% respectivel for testing, and 100%, 91.7%, 95.8% respectively for all confusion matrix. The performance of the ANN classifier increase by increasing the number of samples.

Tumor classification is a vital stage for the performance of the computer aided and automated liver cancer detection. It diminishes the false positive rate by reducing the unnecessary health care cost. Different classifiers were used in medical imaging application like liver cancer detection from CT images however ANN shows very good performance in medical diagnostic systems. In this thesis, an artificial neural network of three layers was proposed for tumor classification. The performance of the proposed structure were evaluated in terms of sensitivity, specificity and accuracy. It was evaluated on 24 images containing 12 benign and 12 malignant tumors. The algorithm works properly in all cases and the structure was evaluated by expert radiologist from preliminary enhancement up to final classification steps.

Ethics: The study was approved by the Ethics Committee of the Tikur Anbessa General Hospital, Department of Radiology, Addis Ababa, Ethiopia.

Acknowledgement

We thank the Center of Radiology diagnosis staff members who gave us samples of Abdominal CT images and their technical supports in the Black Lion General Hospital of Addis Abeba, Ethiopia. We also thank to the Research Square Company for providing automated editing service to our manuscript.

Competing interests: The authors declare no competing interests.

Tarini, S., Tumors: Benign and Malignant. Canc.Therapy and Oncol. Int.J., men’s University, Jaipur, Rajasthan. 2018; 10(3).
Rikabi A., Benders, Al-Marri A., & Al-Thani S., Hepatitis B & C viral infections chronic Liver disease: A population based study in Qatar. Eastern Mediterranean Health J. 2009; 15(4):22–23.
Ernst, A., Image riedufoie desvoies biliaire set du pancreas. Masson, Image Medical Diagnostic. 2005; 20–30.
Clark, H.P., Forrest, W., Carson, P., V.K., Perry S., and Ronald J. Z., Staging and Current Treatment of Hepatocellular Carcinoma. Radiographic. 2005; 25:3–23.
Helenon, Image riedes tumors dufoie. Elsevier. 2003;10,20,25–33,58,131.
Ethiopian Journal Health Development, chronic Liver cancer report. Addis Ababa, Ethiopia. 2017; 31(1):44–56.
Hopper, K.D., Singapuri, K., and Finkel, A., CT and oncologic imaging, Radiology. 2000; 215: 27–40.
Baazaoui, A., W. Barhoumi, A. Ahmad, and Zagrouba, E., Semi-automatic segregation of single and multiple tumors in liver in CT images. IRBM. 2017; 38(2):98–108.
Sahoo, K., S.Solitani, and Wong, A.C., A survey of thresholding techniques Comput. Vision Graph.Image process.1988;41:233–260.
Malik, J., Belongie, S., Shi, J., and Leung, T.Textons,Contours regions: Cue integration in image segmentation. In Proc. IEEE Intl. Conf. Computer Vision. 1999; 2:918–925.
Masutani, Y., Uozumi, K., Akahane, M., & Ohtomo, K., Liver CT image processing:An introduction of technical elements. European Journal of Radiology. 2006; 58:246–251.
Shannon,C.E.,Communication in presence of noise: Proc.Inst. Radio Eng. 1949; 37:10–21.
Al Mahmud, M.A., Karim, A.Z., & Rahman, M.M. Brain. Tumor Detection using MR Images through Pixel based Methodology. Global Journal of Computer Science and Technology. 2015; 15(4):302–315.
Gonzalez, R.C. Woods and Eddins S.L., Digital image processing using Matlab, 2009.

Supplementaryfile.docx
Supplementary

Download PDF

Version 1

posted

You are reading this latest preprint version

Detection and classification of liver cancers using computed tomography images

Status:

Version 1

Abstract

Objectives

Method

Result

Conclusions

Figures

Introduction

Methodology

Image Acquisition

Liver Segmentation

Texture Feature Extraction

Classification of Liver

Liver Tumor Segmentation

Morphological Feature Extraction

Liver Tumor Classification

Analysis of the Classifier

Results

Pre-Processing Results

Liver Segmentation Result

Texture Based Liver Classification Result

Tumor Segmentation Result

Morphological Extraction Result

Tumor Classification Result

Performance of Classifier

Discussions

Conclusion

Declarations

Acknowledgement

References

Supplementary Files

Status:

Version 1