Volumetric Segmentation of Brain Tumors Using the 3D U-Net Architecture
No Thumbnail Available
Date
2022-09
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Addis Ababa University
Abstract
Brain tumor segmentation is one of the most challenging types of medical image segmentations. To
overcome the problems of automated brain tumor classification, a deep learning approach is proposed
herein based on a 3D U-Net model. A 3D model is chosen to get the 3D context of the tumors which are
irregular in shape and could occur anywhere in the brain. In the process of building this model, first, the
data is visualized using different formats of visualization in order to understand the underlying patterns of
the data well. Then the 3D model is developed herein which consists of 22 layers of convolution of which
about half do downsampling and the rest perform an up-sampling of the feature maps. The skip
connections provide more context to the up-sampling layers which change the output back to its original
size. The segmentation model is trained and evaluated on the BRATS 2020 dataset.
There are three versions of the model that are run and observed herein. Of the three versions the second
one, Model2, appears to perform the best. This is the model having data augmentation with batch size of
one. The presented model (Model2) attained a dice similarity coefficient (DSC) of 0.90. The result
obtained shows that the presented method does a good job of segmentation and compared to other state-ofthe-
art methods our technique can be considered competitive in the area of automatic brain tumor
segmentation. In addition, the proposed model is simple when compared to other brain tumor segmentation
models which are complex; this inherently helps it require less segmentation time than earlier models as it
only takes 3.44 seconds to segment one patient’s MRI scan. This contributes to attaining functional realtime
applicable models which can be of real help to physicians and radiologists during the classification of
brain tumor precisely.
Description
Keywords
U-Nets, brain tumor segmentation, dice similarity coefficient (DSC), BRATS, MICCAI