1 minute read

AUTHORS

John Ray Martinez (jbm332@drexel.edu), Jonathan Musni (jem472@drexel.edu), Marvin Joseph Occeño (mr048@drexel.edu), Edmar Parreño (erp75@drexel.edu), and Juan Miguel Trinidad (jbt46@drexel.edu)

This capstone project was selected for oral research presentation at the 2021 Drexel Emerging Graduate Scholars Conference

Snapshots of the comparisons between ground truth, U-Net and PSPNet brain tumor segmentation (Fext = 0.35).

Abstract. Segmentation is the process of examining brain images such as magnetic resonance imaging (MRI) images and computed tomography (CT) scans to locate regions of interests (ROI). These regions define the boundaries of the brain tumor. Undergoing this process allows radiologist or medical personnel to distinguish between healthy cells and tumors. However, with manual segmentation, radiologists take a lot of time and labor to properly segment the images with high accuracy. Human error is also inevitable. With limitations in human capacity, manual segmentation can inhibit diagnosis and therefore delay treatment. Convolution neural network (CNN) models are popular methods in image processing and have rapidly developed as a powerful tool in computer vision and pattern recognition. U-Net architecture is a well-known variant of CNN that consists of contracting (convolution) path and expanding (deconvolution) path to be trained in an image-segmentation map. By utilizing the BraTS 2020 (Brain Tumor Segmentation 2020) from the University of Pennsylvania Center for Biomedical Image Computing & Analytics (CBICA), we are able to investigate the preponderance of this specific architecture. We show that, through hyperparameter tuning of kernel size from 3 to 2 in deconvolution path, sensitivity of Necrotic region has significant improvement while a poor performance in Enhancing region. In this study, we investigate Convolutional Neural Network (CNN)-based architectures such as U-Net and PSPNet. It is found that U-Net outperforms the PSPNet in correctly segmenting the brain tumors. This study’s findings can inform the design and development of an automatic brain tumor segmentation system.

Full paper can be requested.