Abstract:
Brain tumor is the abnormal growth of cells in a brain and they can be benign
and malignant, and malignant ones are considered to be cancerous. There are over
120 types of brain tumor regarding its shape, size and region of tumor [15]. The
common adult tumors are meningioma, glioblastoma multiforme, hemangioblastoma,
pituary adenoma, schwannoma, oligodendroglioma and etc. The most common child
tumors are pinealoma, pilocytic astrocytoma, medulloblastoma, ependymoma and
etc. The brain tumor symptoms vary depending on what part of the brain is affected
and might include headache, nausea, vomiting, sensory disturbances, loss of balance,
seizures [10]. According to how fast these cells grow they are graded from 1 to 4.
For example, the average If 1 and 2 grades mean that they can be treated and cells
grow slowly, whereas grade 3 and 4 brain tumors are malicious and grow fast, and
grow back even after treatment. Mostly, this disease occurs across older adults, but
it can affect people of any age. Brain tumor is complex and hard to detect even for
the radiologists, especially at the early stages. Not all countries have doctors who
can detect brain tumor. Average survival time of people having brain tumor is from
1 to 5 years, according to the stage of brain tumor development. If it is not detected
early, they can grow from grade 1 to 4 and lead to fatal cases. Detection before
grade 1 or 2 brain tumors can help to provide proper medical treatment and prevent
from further spread of disease to other parts of body [15]. MRI is the cornerstone for
brain tumor detection. MRI scanner uses strong radio waves and captures detailed
pictures of brain from different angles, that’s why in this work we will use MRI scans
with corresponding masks, which show the region of brain tumor, to feed the deep
learning model called U-Net. It consists of down-sampling and up-sampling layers,
and will be used for 3D MR image brain tumor segmentation for finding the region
of abnormalities in the brain tissue. 3D MRI scans were in volumetric format and in
4 sequences: FLAIR, T1, T1gd, T2-weighted. The T1 sequence were left out, and
only 3 sequences were used to construct 4D tensors and feed them to 2D and 3D
U-Net deep learning semantic segmentation models. We extracted prediction masks
and compared with the ground truth masks, which were prepared with expert radiologists
for BraTS 2020 dataset. The dice and mean intersection over union scores were used for evaluation. The dice score during training the 2D U-Net reached the
0.8692, and on validation set it reached 0.2465 after 500 epochs. This possibly due to
having unbalanced distribution of data between train and validation sets, therefore
the data should be evenly shuffled in the future. Also, the background label in every
MRI scan had more proportions than the actual tumor part. Therefore, it was hard
for the model to classify to 4 classes, where the 1 class had more proportions, so it
can outweigh the other 3 classes which correspond to the tumor label. The source
code for this work can be found by this link[2]. The work interests focuses on the
development and application of deep learning algorithms and computational processing
in brain tumor imaging, with the aim of improving the assessment, detection and
diagnosis of brain tumor at the time of identification. This work describes the data
preparation, feature extraction, analysis parts in detail and applies state-of-the art
3D U-Net model for volumetric brain MRI scans. The ultimate goal of this work is to
contribute towards making the brain tumor diagnostics decisions more fast, accurate,
and objective.