| dc.description.abstract | Medical Imaging solution for brain tumor diagnosis and treatment consultancy is still
a concerning issue regarding the rapid improvement of technological methods, Artifcial
Intelligence (AI), and its application. Given the current number of brain cancer patients
in need of fast yet precise treatment and the diffculty in transportation to hospital during
recent pandemic years, studies on this problem would be a pragmatic approach and further
be expandable to remote diagnosis.
This study focuses on three aspects. Firstly, I inspect state-of-the-art studies and the
dataset relating to this topic. After choosing BraTS2021 as the dataset, medical image
processing is applied to help refne the input for deep learning models. Secondly, the
Multiscale-GAN model is proposed to perform automatic brain tumor segmentation using
the GAN (Generative Adversarial Network) framework and multiscale learning. Finally,
I integrate the proposed model into our visualization application to provide specialists
with a means to perform tumor segmentation from 3D MRIs within a click.
Compared with fve brain tumor segmentation models, namely U-Net, V-Net, Voxel
GAN, Vox2Vox, and Segtran, our proposed model achieves competitive results with the
highest Dice score of all classes and top outcomes for tumor core class. Additionally,
while having roughly similar results to Segtran, our model is about 17 times smaller than
Segtran. Furthermore, it takes roughly 20 seconds for the model to perform segmentation
using Nvidia RTX 3060 GPU and roughly 3 minutes using Intel i5-4210U CPU. | en_US |
