DL AFUnet 3D BrainSeg

Attention-based Fractal Unet (AFUnet) for brain tumor segmentation using 2D neuroimaging and 3D volumetric brain data

Table of Contents

• Overview
• Code
• Members
• Acknowledgments

Overview

The proposed Attention-based Fractal Unet (AFUnet) combines the ideas of the fractal convolutional network, attention blocks, and autoencoding-decoding Unet structures.

Methodology

Dataset

The dataset we used for 2D segmentation is the The Cancer Genome Atlas Low-Grade Glioma (TCGA-LGG) dataset, which contains 3929 three-channel RGB brain MR images and the corresponding 0-1 tumor mask indicating the regions of abnormalities (label 1's).

The dataset we used for the 3D task is the BRATS-2021 brain image dataset. The dataset consists of 1250 3D volumes with each in the size of $4 \times 240 \times 240 \times 155$. Each sample is composed of four modalities of brain MRI scans, namely native T1-weighted (T1), post-contrast T1-weighted (T1ce), T2-weighted (T2) and Fluid Attenuated Inversion Recovery (FLAIR). The labels contain 4 classes: background (label 0), necrotic and non-enhancing tumor (label 1), peritumoral edema (label 2) and GD-enhancing tumor (label 4).

Code

Please find in folder AFUnet.

Members

Acknowledgments

The authors are indebted to the open-science community for providing resources, knowledge and expertise that serves as a solid foundation for new developments. Also colleagues from Statistics Online Computational Resource (SOCR) and students from the Michigan Multidisciplinary Design Program (MDP) provided support, ideas, and encouragement.

This work was supported in part by the National Science Foundation (grant numbers 1916425, 1734853, 1636840, 1416953, 0716055 and 1023115) and the National Institutes of Health (grant numbers UL1 TR002240, R01 CA233487, R01 MH121079, R01 MH126137, and T32 GM141746)