Multitask Learning for Brain Tumor Segmentation using Transformer
Accurate brain tumor segmentation in MRI images is crucial for effective diagnosis and treatment planning. However, traditional U-Net architecture faces challenges in capturing long-range dependencies and preserving features of small-sized tumors, which limits its performance. In this work, we present MTSegFormer, a novel learning framework for 2D brain tumor segmentation using latent transformer through the Multi-task learning paradigm. We use a UNet-like structure with a latent space transformer, and a self-supervised image decoder to build up the overall framework. We also introduce the Breath-wise Cross Attention module that aims to refine the skip connection features. Experiment shows our proposed framework achieves superior performance compared to other baselines by up to 11% in Dice and 10% in IoU score. The code is available at this url
If you want to use the code, or find our project useful, you can set as follows:
@misc{zhou23mtseg,
author = {Zhou, Meng and Liu, Xudong and Wu, Reyna},
title = {MTSegFormer: A Multitask Learning Approach for Brain Tumor Segmentation using Transformer},
year = {2023},
howpublished = {\url{https://simonzhou86.github.io/portfolio/mtsegformer}},
}