This research work basically highlights my undergrad thesis works. In my thesis, I have worked on the BraTS 2020 dataset. My total journey of thesis from building various models to writing paper is presented here.
The edematous/invaded tissue surrounding the tumor, the necrotic core (filled with fluid), and the enhancing and non-enhancing tumor (solid) core. T1-weighted, postcontrast T1-weighted, and Fluid-Attenuated Inversion Recovery(FLAIR) MRI modalities are extensively utilized to identify the diagnosis, therapy, and evaluation of the disease because they reflect the diverse biological characteristics of the tumor. These MRI techniques facilitate tumor analysis, but require the laborious and timeconsuming manual identification of tumor regions. As a result of deep learning models in computer vision, approaches for the autonomous segmentation of tumor regions have emerged. In an effort to automate the process of segmenting tumor regions, a novel segmentation framework including efficient, contemporary deep learning blocks is provided. Our proposed model is a cascaded encoder-decoder network with two stages. In both stages of training, a variational autoencoder branch is included. In addition, a transformer module is incorporated into the bottleneck layer to account for long-range dependencies. Attention gate is incorporated in the second stage to assist the network in segmenting smaller tumor patches. This block increases the dice score for smaller sub-regions of glioma, such as the tumor that is enhancing. Ultimately, our suggested technique is validated using the BRATS-2020 benchmark dataset. Our method yields equivalent results in comparison to the standard methods. Specifically, 87.09 percent, 80.32 percent, and 74.63 percent dice scores are obtained when segmenting the entire tumor (WT), tumor core (TC), and enhanced tumor (ET), respectively. Ablation study is also undertaken to better comprehend the generalization of the design.
