Brain Tumor Detection Using Multipath Convolution Neural Network (CNN)

Introduction And Background Research

Medical image processing covers wide research scope in the field of computer-aided diagnosis (CAD) for diagnosing various diseases. The analysis of the diagnosis deals with several medical imaging modalities such as MRI, X-ray, radiographic image, CT, mammograms etc. (Sinha, 2014). Image segmentation plays a very important role in the diagnosis and detection of abnormalities in medical images like tumors. There are numerous segmentation methods which are classified on the basis of pixel, region and threshold, referred as pixel based method, region based and object based methods. Segmentation plays very important role in CAD based tumor detection and used just before post-processing. The post-processing determines cancer stage or size and dimension of tumor (Sinha, 2014). Therefore, appropriate soft computing method is necessary to produce most efficient segmentation results. Currently, deep learning methods are used in such applications of medical imaging (Sinha, 2014 & Sinha, 2018). Assessment of human brain and its ability is also investigated with the help of medical imaging data and deep learning (Sinha, 2018) and so as the tumor detection.

Other necessary components of medical image processing employing optimization and deep learning based soft computing methods is de-noising of images because the noisy images if subjected to segmentation then the result and diagnosis analysis would not be appropriate (Bhonsle, 2018 & Sinha, 2017). The k-means method is considered as descent method in segmentation of medical images especially brain tumor images (Patel, 2014 & Patel, 2010) but these methods are also optimized using fuzzy based approaches such as fuzzy based clustering (Sinha, 2015). In (Patel, 2010 & Sinha, 2015), mammograms were well tested with fuzzy method of clustering used for breast cancer detection. Content based medical image retrieval (Singh, 2010), quality assessment of medical images subjected to compression (Kumar, 2011) and application of an appropriate method of optimization in medical image segmentation (Sinha, 2020) are few of important applications of medical imaging and interpretation of images.

In (Neethu, 2017), convolutional neural network (CNN) was implemented for brain tumor detection in which the classification stage employed the CNN based deep leaning method for brain tumor detection. This paper claims to have achieved satisfactory results on the basis of evaluation of the method in terms of sensitivity and specificity though the validation of the evaluation is not reported in the work (Neethu, 2017). In another research (Mohammad, 2016), medical image segmentation uses deep learning tested over 2013 BRATS set dataset as reported and the deep neural network is implemented over MR images of the brain. The method helps in getting contextual features of images much easier and efficiently. The number of epochs that was considered is 10 and as evaluation metrics sensitivity and specificity were presented. This work actually compares a number of CNN architectures that were tested in the medical image segmentation and suggested that these architectures can also be used in brain tumor segmentation and detection in future research (Mohammad, 2016). In (Kazi, 2017), segmentation of medical images was presented using k-means method but the stages of implementation do not cover any role of deep learning (Kazi, 2017).

CNN is tested for segmentation of brain images (OASIS dataset, reported by (Selvathi, 2018) and connectedness and shaped based features are claimed to be obtained in easy going way using CNN and that enhanced the segmentation results as compared to ordinary methods without using deep neural network methods such as CNN (Selvathi, 2018). The method uses only 30 mages in the process of training and test included even less number of images of the said datasets, which is only 10 images. The region segmentation is achieved and the metrics include sensitivity and accuracy in addition to PSNR also. This research tested well with brain images and regions are well segmented but the major limitations include very less number of images tested and the metrics by which the method is evaluated do not show much satisfactory values (Selvathi, 2018). The work is also not compared with any existing method although the CNN architecture suggested by (Mohammad, 2016) is discussed in related research.