A unified patch based method for brain tumor detection using features fusion
Introduction
Brain tumor is the consequence of the production of abnormal cells. It is a substantial reason in increasing mortality rate among adults and children (Amin et al., 2017, Rajinikanth et al., 2017). Based on tumor malignancy and growth rate, brain tumor is divided into two major categories named as primary and secondary. Main tumor initiates from surrounding or inside the brain while secondary level tumor is caused when abnormal cells grow from the primary tumor in other body parts. Based on the origin and behavior of cells, brain tumor is classified by world health organization (WHO) into benign (least dangerous) and malignant (more dangerous). Glioma is a known kind of brain tumor which instigates from the glial cells and infiltrates around the brain tissues (Louis et al., 2007).
A statistical report (http://www.cbtrus.org/factsheet/factsheet.html, 2018) illustrates that overall occurrence frequency of benign, malignant and central nervous system (CNS) tumor is 23.03 cases per 100,000 in 392,982 total incident tumor (7.12 cases per 100,000 malignant tumor in 121,277 total incident tumor as well as 15.91 cases per 100,000 benign tumor in 271,705 tumor incident). The occurrence rate of tumor is higher in females (25.31 cases per 100,000 in 227,834 tumor incident) as compared with males (20.59 cases per 100,000 in 165,148 total incident tumor).
MRI is a painless and non-invasive test that creates images with detailed information about brain and brain stem. MRI images are produced by using radio waves and magnetic field. This scan is also called cranial or brain MRI (Nida, Sharif, Khan, Yasmin, & Fernandes, 2016). MRI is effective in highlighting the brain abnormalities as compared with other modalities because it detects the small structure in brain (shown in Fig. 1). It performs scanning without moving the patient physically (Bullmore and Sporns, 2009, DeAngelis, 2001).
The manual diagnosis process is a hectic and time-taking procedure which degrades the ability of radiologists. Multi-grade detection of brain tumor is a necessary job for better investigation of MRI (Acharya et al., 2019) but still several challenges existwhich reduce the accuracy of tumor segmentation and classification (Rajinikanth, Fernandes, Bhushan, & Sunder, 2018). The major challneges which are faced by researchers in the area of computer vision (CV) are noise in MRI scans, irragular tumors and selection of relevent and irrelevant features (Havaei et al., 2017, Mughal et al., 2018, Rewari, 2019, Sehgal et al., 2016, Swathi and Balasubramanian, 2016). The major contribution of this work is as below:
- I.
Fast Non-Local Mean (FNLM) method is used as a pre-processing step in which 4 × 4 patch and 2 × 2 window size is selected to aid accurate segmentation.
- II.
Otsu algorithm is used for segmentation and applied morphological operations to refine tumor region.
- III.
GEO, LBP and HOG features are extracted and fused in terms of mean and variance to a single feature vector for good classification results. The detailed experiments are performed to compare the results of fused vector as well as individual vector with seven classifiers.
This manuscript is divided into the sections such as existing work in II, proposed approach in III; results in IV and conclusion in V are presented.
Section snippets
Related work
The most relevant work published regarding brain tumor detection using MRI (Amin et al., 2018, Masood et al., 2015, Raza et al., 2012) is discussed briefly. The wavelet (Zhang, Fadili, & Starck, 2008) and combination of adaptive, wiener, median and weighted median filters (Jaya et al., 2009, Krajsek and Mester, 2006) is utilized for the reduction of noise (Greenspan, Ruf, & Goldberger, 2006). The modified curvature diffusion equation (Yang & Huang, 2006), parametric bias field correction, N4ITK
Motivation
We have used the implementation presented by (Amin et al., 2017) for early brain tumor detection in our research. Our research work is greatly inspired by the approach presented for early brain tumor detection in (Amin et al., 2017) where segmentation of brain tumor is performed using Gaussian filter and morphological operations and investigated for shape, texture, and intensity features with variants of SVM (Cristianini and Shawe-Taylor, 2000) classifiers. The authors achieved promising
Proposed methodology
The four major steps (enhancement, segmentation, feature acquisition and prediction) are exploited for the recognition of brain tumor such as FNLM for enhancement, Otsu (Otsu, 1979) algorithm for segmentation, features fusion and classifiers for classification as depicted in Fig. 2.
Material and experimentation
Performance is evaluated on three publicly available datasets, BRATS 2013, BRATS 2015 (Menze et al., 2015) and Harvard (Summers, 2003). BRATS 2013 consists of 30 subjects of glioma, twenty high grade (HG) and ten low grade (LG) while BRATS 2015 consists of 220 subjects of HG, 54 LG and 110 both HG and LG glioma. The Harvard consists of T2 weighted MRI with 39LG and 46 HG glioma images. For validation of this work, two tests are performed, 1st is pixel based while 2nd is based classification on
Conclusion and discussion
Segmenting brain tumor manually is a tedious work. Accurate and efficient detection of early stage brain tumor is the major theme of our work so that the patient could be treated in time before any complication. This method performs well because of the combination of FNLM for enhancement, Ostu for segmentation, features fusion and multiple classifiers.
The proposed technique achieved 0.98, 1.00 and 1.00 ACC on BRATS 2013, 2015 and Harvard datasets respectively. The comparison of results with
Declaration of Competing Interest
The authors declared that there is no conflict of interest.
References (78)
- et al.
Detection of brain tumor in 3D MRI images using local binary patterns and histogram orientation gradient
Neurocomputing
(2017) - et al.
A novel machine learning approach for scene text extraction
Future Generation Computer Systems
(2018) - et al.
A hybrid feature extraction approach for brain MRI classification based on Bag-of-words
Biomedical Signal Processing and Control
(2019) - et al.
MRI segmentation fusion for brain tumor detection
Information Fusion
(2017) - et al.
A mix-pooling CNN architecture with FCRF for brain tumor segmentation
Journal of Visual Communication and Image Representation
(2019) - et al.
Computer-aided diagnosis of human brain tumor through MRI: A survey and a new algorithm
Expert Systems with Applications
(2014) - et al.
State of the art survey on MRI brain tumor segmentation
Magnetic Resonance Imaging
(2013) - et al.
Brain tumor segmentation with deep neural networks
Medical Image Analysis
(2017) - et al.
Efficient multi-scale 3D CNN with fully connected CRF for accurate brain lesion segmentation
Medical Image Analysis
(2017) Glioma detection based on multi-fractal features of segmented brain MRI by particle swarm optimization techniques
Biomedical Signal Processing and Control
(2017)