Efficient nnU-Net for Brain Tumor Segmentation

Abstract

Brain tumors are one of the leading causes of death in adults. They come in various shapes and sizes from one patient to another. Sometimes, they infiltrate surrounding normal tissues, making it challenging to delineate tumor boundaries. Despite extensive research, the prognosis is still low. Accurate and timely brain tumor segmentation is critical for treatment planning and disease progression monitoring. Automatic segmentation of brain tumors using deep learning methods has produced high-quality and reproducible segmentation results. Specifically, the encoder-decoder networks, like the U-Nets, have dominated the previous BraTS Challenges because of their superior performance. Due to the importance of high-quality segmentation, most state-of-the-art models focus more on pushing the boundaries of the current methods at the expense of computational complexity. The computational budget for practical applications is minimal, requiring technological solutions that balance accuracy and available computational resources. In this study, we extended the U-Net model in the nnU-Net by replacing the basic 3D convolution blocks with bottleneck units utilizing depthwise-separable convolutions. Furthermore, we introduced the shuffle attention mechanism in the skip connections to compensate for the slight loss in segmentation accuracy due to a reduction in the number of parameters. On the brain tumor dataset BraTS 2020, our network achieves dice scores of 79.2%, 91.2%, and 84.8% for enhancing tumor (ET), whole tumor (WT), and tumor core (TC), respectively, with only 2.51M parameters and 55.26G FLOPS. Extensive experimental results of the BraTS 2020 dataset reviewed that the proposed modifications achieved competitive performance at a lower computational cost. The code for this project is available at https://github.com/tmagadza/EfficientNNUNET.git.

Keywords

• Brain tumor segmentation
• depthwise-separable convolutions
• group convolution
• shuffle attention
• U-Net