A Survey on Digital Image Processing Techniques for Tumor Detection

The paper presents a formal review on evolution of the image processing techniques for tumor detection, comparison of the existing techniques to obtain the one which gives the best results for detection and classification of tumor. The scope of the propounded technique is semblance of the gaps by giving effectual results in identifying the tumor.


Introduction
The paper aims to make a comparative analysis on different tumor detection techniques, and results are made on the basis of parameters considered, so as to find the robust algorithm for tumor detection.
The origination of the term cancer in (460-370 BC), is credited to the "Father of Medicine", Greek Physician Hippocrates. However, having its earliest evidences as osteosarcoma (bone tumors) in fossilized human mummies in Ancient Egypt (3000 BC), cancer, in accordance with latest statistics is amongst the leading causes of deaths all over the world as it's a life threatening disease.
Cancer in living beings occur when the DNA, basis of genetic code of cell gets corrupted due to an exposure to chemicals, radiations, inheritance or viruses that lead to mutation in the genetics.
The 19 th century was the birth of scientific oncology when Digital Images began to be used for screening and early detection of tumor and its classification into malignant or non -malignant as shown in Table 1. The concept of the images, that aroused millions of years back has its roots in nature because of its origination since the existence of a source of light, today has become of great help to the experts in detection and accurate location of tumor sites in humans which was earlier difficult due to complex pathologies.
The scope of this review is to address various image processing techniques being used for tumor detection and considering there pros and cons. On the basis of their pros and limitations it becomes relatively easy to detect tumor at a much early stage. From the literature its ascertained that the malignant tumor bear out to be dangerous and can certainly lead to death. Meticulous attempt has been made in the paper to find the robust and vigorous technique that can precisely classify the tumor as malignant or benign

History Specifying Relation between Tumor Detection and Image Processing
The baleful disease tumor has evolved since human beings existed. Since 2500 BC, its detection is considered a vital subject. The following table describes how tumor detection techniques have evolved over time since it existed. The techniques being used from earliest time till present era with their benefits, limitations and accuracy are discussed in Table 2.

Tumor Detection Process
In this section complete and the basic procedure of tumor detection by Image Processing is described with help of DFDs. In DFD level 1 the intermediate step is presented in which the region of interest is detected, by firstly pre-processing the image and then applying segmentation. After

Feature Extraction
This technique is of great importance in field of Medical Image Processing. When an image of particular body part is to be studied for diagnosis of tumor, feature extraction provides an aid by reducing dimensionality or area to be studied. This approach is also used for segmentation of images, to find ROI. Rather than considering all the image features at once, selection and extraction of good features will lead to better segmentation results. Feature extraction 31 also plays a vital role in classification of tumors. For methods of constructing combinations of the variables feature The secondary tumors are those which originated at some other locations in body but spread to other parts as well.  finding the region of interest, classification is done in order to find out whether the suspected candidate region is a tumor region is tumor or not. DFDs level 0, 1 and 2 are presented. DFD level -0 is a simple representation of detecting tumor region in the acquired image by applying different techniques of Image Processing and Machine Learning. DFD level 1 as in Figure 2 explains the concept behind detection of the region of interest. DFD level 2 as in Figure 3 depicts the complete procedure of tumor detection process, presenting the various techniques being used in the process. The techniques are used individually, or in combination for finding the tumor. The techniques Ds. In DFD level 1 the intermediate step is presented in which the region of interest is detected, by firstly precessing the image and then applying segmentation. After finding the region of interest, classification is done in er to find out whether the suspected candidate region is a tumor region is tumor or not. DFDs level 0, 1 and 2 are sented. DFD level -0 is a simple representation of detecting tumor region in the acquired image by applying ferent techniques of Image Processing and Machine Learning. D level 1 as in Figure 2 explains the concept behind detection of the region of interest. DFD level 2 as in Figure 3 picts the complete procedure of tumor detection process, presenting the various techniques being used in the process. e techniques are used individually, or in combination for finding the tumor. The techniques described below also ve sub techniques, which are used to detect tumor. Some of the techniques are pre-processing, template matching, ural networks etc. per the literature the basic techniques of image processing and machine learning being used for tumor detection are cussed below. Preprocessing e pre-processing operations 1,15,19 basically facilitates subsequent processing of digitized images which get ntaminated with noises during acquisition process and thus image data gets corrupted, but with help of precessing techniques the information of all the distorted pixels can be restored. DFD level 1 as in Figure 2 explains the concept behind detection of the region of interest. DFD level 2 as in Figure 3 depicts the complete procedure of tumor detection process, presenting the various techniques being used in the process. The techniques are used individually, or in combination for finding the tumor. The techniques described below also have sub techniques, which are used to detect tumor. Some of the techniques are pre-processing, template matching, neural networks etc. As per the literature the basic techniques of image processing and machine learning being used for tumor detection are discussed below.

Preprocessing
The pre-processing operations 1,15,19 basically facilitates subsequent processing of digitized images which get contaminated with noises during acquisition process and thus image data gets corrupted, but with help of preprocessing techniques the information of all the distorted pixels can be restored. [*Corresponding Author] Page 4

Feature Extraction
This technique is of great importance in field of Medical Image Processing. When an image of particular body part is to be studied for diagnosis of tumor, feature extraction provides an aid by reducing dimensionality or area to be studied. This approach is also used for segmentation of images, to find ROI.
Rather than considering all the image features at once, selection and extraction of good features will lead to better segmentation results. Feature extraction 31 also plays a vital role in classification of tumors. For methods of constructing combinations of the variables feature extraction is a general term to get around these difficulties while describing the data with adequate precision.

Segmentation
Partitioning of digital images into sets of pixels is called segmentation. It aims to represent an image in a much simpler way, to make it easier to analyze and much more informative. As a result of segmentation set of segments 30 is acquired which covers entire picture collectively. Every pixel within a segment possess same characteristics like color, intensity, texture and label as others. In medical imaging segmentation is of extreme importance to find contour of different anatomical structures.

Segmentation
Partitioning of digital images into sets of pixels is called segmentation. It aims to represent an image in a much simpler way, to make it easier to analyze and much more informative. As a result of segmentation set of segments 30 is acquired which covers entire picture collectively. Every pixel within a segment possess same characteristics like color, intensity, texture and label as others. In medical imaging segmentation is of extreme importance to find contour of different anatomical structures.

Classification
The definitive step in tumor detection is classification of tumor as malignant or benign tumor. Classification 34 in general involves two steps of training and testing.

Template Matching
Template matching is a technique used to verdict regions in image which are similar to template. In tumor detection process, a template is used as an aid for coarsely locating tumor by ruling out the edges. A template is created and is moved over the acquired image sequentially and the area where the templates match the image is marked. The template is always of smaller size than image. The technique is of great importance and is useful in successfully detecting tumor

Neural Networks
The concept of Artificial Neural Networks was originated from "Central Nervous System" of humans. The artificial neurons in ANN reminiscent of human neurons are basically interconnected nodes. This concept in practice since 1943 works according to the training and learning process, is of the essence in field of medical imaging and is used for enhancement of images and classification of tumors. It has helped to make the tumor detection process automated

An Overview of Some of the Existing Techniques for Tumor Detection
As we know from history, that tumor is life threatening for humans, therefore, its early detection and treatment is necessary as it would increase the survival chances of the patient. Presently, many imaging modalities exist which are being used by radiologist for visualization of the internal structure for detecting various diseases without performing surgery. According to the latest statistics there are around 14.1million cases of cancer around the world every year. This poses the biggest challenge to the radiologists who manually locate the lesions on different modalities. Thus, an accurate computer aided clinical tool is an imperative necessity in order to assist the experts detecting tumor. As mentioned in Table 3 above there are various techniques which were used to detect tumor since 3000 B.C. Some of the techniques were very effective, while certain procedures led to the need of developing robust techniques.
Over the centuries there was a great change in beliefs towards the disease which had a major impact on the advancement of techniques for detecting tumor. Years later in 19 th century images began to be used to diagnose tumor In 1 proposed an approach to augment the segmentation for extracting suspicious mass areas from mammograms. The paper focused on automatic detection of all possible lesion sites by using Enhancement and model selection, A morphological operation "Dual Morphological Operation". Histogram models, FGGM distribution to determine the Kernel shape and number of regions in image.
In 4 presented work on techniques for detecting circumscribed masses in mammograms. In the initial step mammograms were enhanced by removing noise in background while preserving the edges.
In 3 presented a review on all the components of image analysis system which includes four steps -Pre-processing, Feature Extraction, Segmentation and Classification, used to MR detect Glioblastoma and evaluated the results in all the cases as shown below in Table 4.
In 5 proposed a method to detect masses in mammograms and classifying them as benign and malignant by using texture field analysis. Success rate of 100% was achieved.
The pre-processing techniques used for enhancement were Gaussian low pass filter and Pyramidal Decomposition, which helped to smoothen the image.
In 6 proposed an algorithm for unsupervised color segmentation with appliance to skin tumor.
By using color feature to detect tumor border, author put forward an algorithm 5 which involved noise removal with Pseudo Median Filter and non skin masking algorithm.
The Table 5 presents all the algorithms used with their average error and standard deviation.
In 7 proposed a filter for detection of lung nodule on chest X-Rays, three kinds of CI filters had been investigated and experimentally evaluated. Different characteristics of CI Filters Adaptive Ring, Coin and Iris were discussed in detail .
In 8 proposed a technique to detect tumor by measuring abnormal thickness in bladder wall by taking the mean and standard deviation. The Table 6 below presents the comparison of automated and manual process.
Thresholding and marching cubes algorithm were used for segmentation, a normal thickness atlas with average and other with Standard deviation was constructed 8 .
In 9 presented a CT liver image diagnostic system to unearth the liver boundary and tumors in liver. To extract the liver boundary segmentation approach "Detect Before Extract" was used. The desired liver boundary was achieved by using Deformable Active Model (contour modification model). Classification system developed on basis of MPNN (Modified Probabilistic Neural Network) was applied for discrimination between heptoma and hemageoma liver tumor. The classification system used in this paper is Texture -Feature based which used features.
In 2 proposed a technique to recognize nodular tumor in chest radiographs by using a Ladder Structured  No.of films tested 17 2 No. of candidate sites 19 4 Average no. of candidates per film 1.1 5 Hit rate 100% 6 False alarm rate 1.7  1) Digitization of radiographs 2) Enhancement or pre-processing operations which results in edge detection. 3) Recognition of tumor and its classification. The technique used gave accurate and comparable results. The algorithm is explained as follows.

Recognition of Tumor and its Classification
Initialize Image data set I SET = 1 to n If I SET(n) = P × P // where P  10 proposed an approach to detect lesion in liver. Median filtering was used to reduce noise in image while contrast information was retained. To determine closed and accurate contours Snakes/Active contour method was used. The following Equations describe the snake model.
Energy minimization splines Snake models are guided by internal and external forces. E internal is comprised of the continuous energy and energy of curvature and the E external is composed of the image energy and other constraints enforced by user.
Equation (2) depicts the discrete form of the snake equation. E Cont represents the energy because discontinuities.
The energy due to discontinuities as follows.
In Equation (3), d av represents the average distance between two snake points which are neighbors.
S liver = (T 1 -G (x,y))/(T L1 -T VL ) Two fuzzy sets corresponding to fuzzy edge region and fuzzy smooth region, S lesion, S liver were constructed using the two new thresholds T V1, T L1 which have equal space between them and the gradient magnitude values using sobel operator.
In 11 proposed an approach for early detection of skin cancer and measuring the vascularisation and pigmentation in Nevoscope images. Pre-processing pace involved resizing, masking, cropping, hair removal and converting RGB to Gray scale images. The different segmentation techniques used were Sigmoid mapping, Fuzzy C Mean , Principle component transform. The best segmented image out of three was selected by considering the correlation coefficient, edge strength and lesion size. Transillumination image was used for classification of tumor as benign or malignant.
In 12 proposed an algorithm for segmenting brain to detect tumor using alignment based features. The preprocessing pipeline was used which included: 1. Non linear filtering for noise reduction, 2. Inter -Slice intensity variation correction, 3. Intra volume intensity bias field correction, 4. Alignment of different modalities 5. Linear alignment of modalities with template, 6. Non linear warping of modalities with template. 7. Re-sampling of voxels to template coordination system ( B -Spline). 8. Weighted regression for inter volume intensity standardization 12 In 13 presented a novel approach in this paper for detecting brain tumor in MR images. Abnormal tissues were detected using co-variance method and (Principle Component Analysis) PCA.
In 14 proposed a technique for characterization of brain tumors. The presence of tumor can be determined by any of PCA, ICA, ISOMAP algorithms. In the pre processing stage enhancement and resizing of 113 3D MR images was done. The position of tumor was manually indicated by experts, median intensity was used for having all intensities in same range. Gentle Adaboost algorithm 14 was used for training the classifier which used intensity based features, gabor features, symmetry features.
Adaboost algorithm gave the best results for classification of tumors as benign and malignant. The results obtained by the algorithm are comparable.
The proposed technique is suitable for part based segmentation.
In 15 proposed a Pre-Operative and Post-Recurrence method for brain tumor registration. PORTR works on facts that all scans are from same patient. Minimum of energy function is used to determine the deformation between two scans. The pre-operative and post recurrence scans of 24 patients were experimentally registered based on discrete optimization, ANTS, Attribute vectors, Mutual saliency. The method gives accurate results the future work can be done to enhance the algorithm for even better results.
In 16 proposed a technique for tumor detection in cervical tissue. Microscopy images were used, Ki-67 was used to provide assistance in tumor detection process. Anisotropic Diffusion was employed for noise removal since it preserved the edges Firstly the abnormal nuclei were detected, then all the touching nuclei were detected and the other regions.
Finally all the tumorous cell were detected as the nuclei of the tumorous cells is always of larger size than the other .
In 17 proposed a speculation detection method, which is of great importance in characterizing malignant tumors. 3D ultrasound was used for detecting speculation encasing breast tumor. Morphological operations and ROSE Algorithm were used for region selection. STICK Algorithm was used for ultrasound segmentation.
After applying the thresholding method, the pixels were separated into regions based upon their intensity difference.
The region which comprised the centre of the image was selected. In the equation above the binary image is represented by I(x,y), (I,j) signify the centre and N is the number of marked pixels.
For the segmentation of the image, gradient of image and orientation structure is calculated. The gradient is given by and can be represented as in Equation (2), and the orientation structure is given by -  In case of the ultrasound images the method for segmentation described above does not work. Therefore, Stick Algorithm was used which is based upon sticks, line segments in different angular orientation.
The corresponding angle and orientation is calculated as in Equation (4) and (5) respectively - q j -(i min -1)π/(2n -2) In 18 used phase contrast images and machine learning approach for detection of tumor cells. Super pixel labeling was used for processing. Simple Linear Iterative Clustering (SLIC), and K-mean clustering was used for segmentation and SIFT for extracting features. CS-LBP was employed for region description. Random Forest Classifier was used for classification.
In 19 put forward an approach to detect ROI, by using morphological band pass filters in digital mammograms.
Multi level morphological operations were used for pre processing of mammograms ROI was detected by applying thresholding method. Binary morphology operators were used for refinement process. The technique gives more accurate results than DWT.
In 20 proposed an algorithm for classification of mammograms grouping them into three categories benign, malignant, normal.

SENSITIVITY = TP / TP + FN
SELECTIVILTY = TP / TP + FN (12) In 21 in their work used a set of 251 digitized images. Automatic induction was used for producing classification Artificial intelligence concept was used for classification of tumors.
The study 21 aspired to prove that different features extracted from colored skin tumors are sufficient to distinguish the benign tumors from malignant.
The major aim was to precisely find the malignancy of the tumor since it is deadly for the patients suffering from cancer.
Results of the algorithm are presented in Table 7, 8 and 9.
In 22 , proposed a method for classification of breast masses by image processing techniques. During the pre processing step, different techniques like Anisotropic Diffusion Filters and Gaussian filter were used for removing noises.    To obtain a closed Boundary of contour of speculated mass, Active Contour Model 22 was used.
In the Final step classification is performed using shadowing features and shape features by Support Vector Machine 31 . The technique is non-invasive and cost effective. The Table 7 below describes the result of the approach.
In 23 proposed an approach for tumor detection in brain by using Texture based region growing and Edge detection techniques. MR images were used. Various techniques Gray level conversion, resizing of image, median filtering, high pass filtering 37 were applied for pre processing. After pre-processing all the noises were removed from the image which gave better results in detection of tumor 23 . Modified texture based region growing was used for segmentation. Moore neighbourhood method was employed for edge detection of cells. The results obtained are presented in Table 8, 9, 10. In 24 proposed an automatic diagnostic system for detecting pulmonary nodules in CT images. The images are pre-processed and enhanced. On the basis of medical knowledge three rules are generated which used the features for classifying the candidate nodules as tumors or non tumor.
In Table 11, comparison has been made between results from physicians and CAD system. The results are compared on the basis of physician in agreement with results from CAD system.
In 26 proposed a new fusion model for classification of lungs tumor by using Genetic Algorithm.
The CT images were pre-processed using median filter 32 and morphological operators.
Texture analysis and a novel fusion MAD technique were used for feature extraction.
Genetic Algorithm was utilized for selecting features and K -Nearest neighbour 34 multilayer preceptron -NN 33 for classification of tumor as benign ormalignant. In Table 12, complete detail of classification of lung tumor. Table 14 performance of different classifier is discussed as emphysema, bronchiestasis, pleural effusion and normal is presented in case of different classifier.
The no. of correctly identified cases are mentioned. In Table 13, accuracy in terms of percentage in case of different feature selection and extraction technique is mentioned.   In 25 applied an Artificial Neural Network for detection of breast tumors detection in mammograms. SOM was used for reducing the Dimensionality. SVM kernel based approach was used for classifying the tumor as benign or malignant. For interaction between various variables MARS was used. Sensitivity, Specificity, Youden index Accuracy were used as performance measures. Clustering algorithm were widely used as exploratory tools for analyzing the breast cancer data. Nevertheless, there are certain limitations in case of clustering algorithm. In, clustering algorithm based on SOM 36 was used for pre-processing the information in the samples of breast screening programme.
The performance measures considered are sensitivity, accuracy, specificity 35 and Youden index. The MARS model, used for modeling as it builds models piecewise by using linear regression. The performance of MARS model is discussed in Table 15 above taking specificity and sensitivity as performance measure. Table 16 and 17 presents the confusion matrix and performance measures of SVM model.
The death rate of cancer patient was too high in earlier times but now with much awareness the mortality rate has decreased as different techniques have been developed for early many technique have been developed since 3000 BC when there was no cure for the baleful disease.
The development of these techniques have lead to decrease in the mortality rate of the cancer patients. Ever since tumor existed the techniques have evolved and have been improved for better results. There are different techniques developed for detecting tumor with varied accuracy and hit rate. Table 18 compares the different techniques based upon their accuracy, sensitivity and specificity. Depending on their accuracy, the suitable algorithm can be easily determined, which will lead to the accurate detection of tumor.
The techniques with less accuracy rate, sensitivity and specificity can be worked upon for enhancement.

Methods and Material
The section presents a detailed review of different techniques being used for tumor detection over decades.     Table 19.
In Table 19 all the techniques been used for detecting tumor between 1970's and 2000 have been discussed. The period between 1995-1998 was a revolutionary era in the field of tumor detection, ample work was done as there was widespread awareness about the disease. Most of the work was driven on the modalities -mammograms, CT scans. The results were comparable but the technique could be improved.
All the techniques were able to detect the tumor but efficiency was low. In the year 1999 there were many techniques akin to the methods developed during 1995-1998 were developed with high success rates. Many novel techniques for detecting tumors were proposed as discussed in this era, but with a need of enhancement as many of the techniques developed did not adhered to time and cost and also were less efficient. Thus initiative can be taken to enhance these techniques. Many techniques with high accuracy rate were presented.
During this era there was an emerging trend towards the development of tumor detection tech-  The technique proposed is strong to detect tumor but can be improved for classification purpose in the field.
The Table 21 presents all the techniques being used between 2005-2015.
The latest techniques being used for detection are discussed with their pros and cons.
The limitations of the techniques are taken into consideration, in order to improve their efficiency by putting more emphasis on them. In 20 proposed an algorithm of detecting the tumor and classifying it by using decision forest classifier. The technique achieved the success rate of 90%.
In 16 used microscopic images to detect the abnormal cells based upon the tumor detection marker and separating their contours. The hit rate of the technique used was 95%. In 25 proposed an algorithm to find tumor in breasts by using mammograms. Different techniques like

Scope of Work
From the extensive literature survey, its observed that an automation system is entailed for detecting and classifying tumor efficiently. The technique, AVG, proposed for detection of tumor is an amalgamate of adaboost and genetic algorithm.

Conclusion
The papers endeavoured accomplishment of a partial survey on diverse tumor detection techniques for different modalities. A comparative analysis is made on various techniques. After evaluating various approaches, the methods which can efficiently and accurately detect tumor are mentioned in results. The paper puts forward a new algorithm which is an amalgam of adaboost and genetic algorithm proposed to focus on accomplishment of more accurate results than existing techniques. The work will be extended for building new algorithm for tumor detection with an aim to reduce the computational time and implementation cost. As detecting tumor is a very complicated and sensitive task, accuracy and reliability are of much importance. Thus, a sophisticated technique that highlights new vistas for developing more vigorous image segmentation technique for tumor detection is much sought after