MRI Brain Tumour Segmentation using a CNN Over a Multi-parametric Feature Extraction

A Brain tumour is a collection of an abnormal mass of tissue that can be grown as cancerous. This pathology can be detected using noninvasive techniques such as CT and MR. Despite CT can form a three-dimensional computer model by taking multiple X-rays shots, the MRI scans are highly preferred since they do not use ionizing energy on its captures and they also provide sufficient information to confirm a diagnosis, however, MRI scans have a lot of noise which can reduce the accuracy of the diagnosis. Therefore, many works in the state of the art try to solve these issue using first a filtering method to clear the noise and then a semantic classification algorithm such feature pyramid network, mask R CNN and random forest classifiers trained over the images acquired with MRI technique extracting grayscale intensity, spatial proximity and texture similarity features, however, segmentation image using these methods does not have sufficient accuracy. Thus, this work proposes to look forward over the FLAIR images on the BRATS 2015 training dataset that is composed by 155 captures of axial cuts from where the principal and adjacent layers that have the highest amount of information are used to reformulate and increase data features that lead on a pixel-based classifier U-net proposed performs a semantic segmentation with a precision of 76%, which improves in up to 23% precision compared with the random forest-based method that obtained a 53% of precision.