An Efficient Computer-Aided Diagnosis System for the Analysis of DICOM Volumetric Images

Medical images are an important source of diagnosis. The brain of human analysis is now an advanced field of research for computer scientists and biomedical physicians. Services provided by the healthcare units usually vary, the quality of treatment provided in the urban and rural generally not same. Unavailability of medical equipment and services can have serious consequences in patient disease diagnosis and treatment. In this context, we developed. MRI (Magnetic Resonance Imaging) based CAD (Computer Aided Diagnosis) system which takes MRI as input and detects abnormal tissues (Tumors). MRI is the safe and well reputed imaging methodology for prediction of tumors. MRI modality assists the medical team in diagnosis and proper treatment plan (Medication/Surgery) of different types of abnormalities in the soft tissues of the human body. This paper proposes a framework for brain cancer detection and classification. The tumor is segmented using a semi-automatic segmentation algorithm in which the threshold values selection for head and cancer regions are premeditated automatically. Segmented tumors are further sectioned into malignant and benign using SVM (Support Vector Machine) classifier. Detailed experimental work indicates that our proposed CAD system achieves higher accuracy for the analysis of brain MRI analysis.

such as development and segmentation of an image [1] to find-out analytic figures like recognition of damaging area or disease. The brain of human analysis is curretly an advanced research field for computer scientists and bio-medical physicians [2]. The brain is the core part of the human CNS (Central Nervous System), it is protected via a strong bone called a skull. The total average mass of a normal human brain is 1300-1500 mg and volume is 1260 cm 3  descriptions and support the radiologists to research on the human brain in depth. Physical analysis of skull MRI is time taking and vulnerable to errors due to the size of MRI data and complexity for every patient. So, many state-of-art segmentation methods have been presented for this type of situation; each technique has both advantages and disadvantages [3]. Most of the work done on brain MRI analysis has been made on a single 2D image, which has been converted from DICOM segmentation of brain and tumor in a single image is easier than the segmentation of volumetric images, but the results obtained are not beneficial for the surgeons and radiologists. The segmentation phase is also an important phase for several dimensions such as calculation of tumor volume in cubic millimeter, 3D therapy 360 0 rotatation, and sorting of cancer type into benign and malignant. Thresholding typed segmentation is a famous technique for the MRI volumetric image segmentation [4]. Working of threshold-type segmentation is depended on the collection of a preferable value of the threshold.
Physical collection for the correct edges is complicated and needs machine learning knowledge. For this kind of operational test support unitbinarized auto-segmented brain and tumor parts with the support of selecting threshold correct values [5]. This Otsu binarized way is cheap to apply but it is noisaters and sustains segmenting cancer having similar existence to other brain portions. In the Watershed [6] segmentation, objects are overlapping because it is a robust technique for segmentation. Watershed applies with the magnitude of gradient on the image to find segment objects of interest or tumor. The overall performance of watershed technique is good, but it is not satisfactory for 3D volumetric classification. Clustering is an unverified knowledge method in the brain knowledge used to combine unbalanced items with matching characteristics in a collection [7]. According to clustering scenario procedures, K-map is recognized as a clustering procedure that is utilized for image processing and collecting based facts [8]. Area increasing [9] known as segmentation technique of image and mounted on optimal seed facts, region growing algorithm can be single-seeded or multi-seeded. Multi-seeded point region growing is better than a single-seeded point because the region grows algorithm is very sensitive to noise. Area increasing algorithm is a partial automatized technique because the seed points are selected manually, so choosing a good kernel edge is a complicated job.
FCM (Fuzzy C Mean) [10] is a vast familiar collecting procedure same as K-map and other FCM conversion is more complicated than K-map image division [11].
In the current research, a hybrid segmentation algorithm is developed. A GUI (Graphical User Interface) component is used to select five seeds from the brain and tumor region. Based on these five seeds automatic threshold values for brain and tumor are calculated.
After image-segmenting the human brain into various neurons, classification is applied to each tissue to analyze whether parts are damaged or not. In this After, classification and segmentation, brain size can be designed by using the material found from DICOM and tumor region.

LITERATURE REVIEW
In this review, both sgmentation and classification are discussed. Image segmentation is the pre-processing operations in a medical image that is done before classification and other post-processing operations including volume calculation and 3D visualization. Zhang et. al. [1] represented a brief study on present methods associated with MRI brain study with deliberating many partition procedures for brain cancer segmentation [17].

METHODOLOGY
The proposed framework involves two stages: (1) Calculation of an optimum threshold for segmentation of brain and cancer sections

Brain Tumor Segmentation from Volumetric MRI Using Global Thresholding Algorithm
MRI is processed for the sake of enhancement, restoration, and extraction of expressive data from the images. MRI uses the 16-bit monochrome DICOM format which is indeed complex and time consuming for processing manually and automatically. Similarly, the brain MRI data not only consists of the brain, but it also contains skull, Cerebrospinal fluid, skin, and fats.
Threshold segmentation is easy and more commonly used in CV (Computer Vision) and image processing application.
The aim of thresholding is to divide a copy into two clusters: one cluster will be the foreground and the other cluster will have background values. The single and Multilevel threshold value is a manual technique and a threshold value calculation for single image 2D MRI slices does not provide satisfactory and meaningful segmentation results. A better approach is to segment 3D volumetric data automatically.

An Efficient Computer-Aided Diagnosis System for the Analysis of DICOM Volumetric Images
Otsu based automatic segmentation is used in CV and image processing to cluster an image in bi-levels or to reduce the gray levels 0-65536 into binary 0-1, 0 represent the background and 1 represent the foreground. This method creates a bi-modal histogram separating the gray levels into two separate classes for the sake of finding an optimal threshold value, the combined spread (intraclass variance) should be minimal. Otsu segmentation computes an optimal threshold value using different statistical parameters (weights, standard deviation, and variance) for the segmentation of interest regions in an image. The voxels are clustered on the basis of optimal threshold value, intensities greater than or equal to the optimal threshold are clustered as foreground voxels and intensities less then optimal threshold are clustered as background voxels.
The algorithm of OTSU is explained using a 6x6 image with only 6 grayscale levels as shown in Fig. 2(a-b).

Calculation for an Optimal Threshold Value
Calculating the foreground and background variance for obtaining an optimal threshold value is show in Fig. 3.
Following steps are explained in [21].

7.
Calculating the Intra Class Variance: MRI volumetric data is represented using the histogram in Fig. 4. Automatic segmentation of brain and tumor regions is shown in Fig. 5.
Morphological operations are performed on the segmented brain and tumor images to remove unwanted areas/voxels from the segmented brain and tumor images.
In these operations, a 2D or 3D small image (kernel) is created which consists of dissimilar kinds of numerical

Features Extraction using Edge Histogram Descriptor
In our proposed framework, we used EHD (Edge Histogram Descriptor) for feature extraction.  7(a). MORPHOLOGICAL ERODED BRAIN FIG. 7(b). MORPHOLOGICAL ERODED TUMOR EHD descriptor compute the direction of edges in different angles. In the first phase of EHD descriptor, the image is divided into 16 non-overlapping blocks of the same size.
Each image block is then split into four sub-blocks and assigned the labels from 0-3. Average gray level for each block is then calculated as a k0 (i,j). For each direction, filter co-efficient is represented as f. Each of the image blocks is then classified into one of the five mentioned edge categories or as a non-directional edge block. Now, the magnitude is represented as m. The proposed framework is shown in Figs. 8-9(a-e).

SVM classifier is used for classification, Supervised
Classifier is trained on the edges angles features extracted from the ROI (benign and Malignant). The proposed classification framework is shown in Fig. 11.
Then the classifier finds out the difference between

Evaluation of the Proposed Segmentation Method
A subjective performance evaluation of the proposed segmentation method is performed using sensitivity,

Dataset and Tool
The