Abstract:
Brain tumors are a common type of gliomas that are caused by abnormal spread of brain cells. They can be malignant or benign; and depending upon their type, the fatality rate varies. Hence, early detection of brain tumors helps in finding the best treatment response to increase the rate of survival of the patient. However, segmenting out the brain tumor manually from a huge amount of dataset of images that are generated in clinical trials, is not only difficult but also consumes a lot of time. Brain tumor segmentation is considered a complex problem owing to the variability of the shape of the tumor and difficulty in determining its location. Due to this, there is a need for automated segmentation of unhealthy brain tissues. Although a lot of areas are being explored to reach the best solution, most of the algorithms of deep learning being developed are very complex. Deep learning gives required results with a good accuracy but simple procedures have yet to be explored in detail for such complex problems. Hence, the aim of this research is to show the relevance of using a simple k-nearest neighbor (knn) classifier to solve the complex problem of segmentation in multimodal magnetic resonance images (MRI). The dataset utilized in this study for proposed segmentation algorithm, is of BRATS challenge 2018. The algorithm was trained on the provided training dataset and was later validated using a different dataset for validation that was provided. Training data consisted of High-Grade Glioma (HGG) patients and Low-Grade Glioma (LGG) patients of which 40 of each has been used so far for the training. This research shows that the complex problems, such as this one, can be approached using a simple technique. It can suffice for the detection of the tumor as well as segmentation. However, their accuracy will be lower compared to the deep learning algorithms. However, in this algorithm that has been designed, knn is just the basic step to label the MRI slices containing tumor. Further mathematical analysis was done on the dataset to find the location and progression of the tumor. The overall accuracy of the algorithm to precisely classify the tumorous slices is 80.76% with specificity and sensitivity 0.84519 and 0.75159 respectively.
