Improved U-Net3+ with stage residual for brain tumor segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• The U-Net3+ network exhibits insufficient brain tumor feature extraction, leading to poor feature fusion and reduced segmentation accuracy.
• Deep convolutional neural networks face challenges with vanishing gradients as network depth increases, hindering effective feature learning.

Purpose of the Study:

• To enhance the feature extraction capabilities of the U-Net3+ encoder for improved brain tumor segmentation.
• To develop a novel segmentation network that addresses the limitations of existing models, particularly in handling 3D medical data.

Main Methods:

• An improved U-Net3+ network incorporating a stage residual structure in the encoder to enhance feature extraction and mitigate vanishing gradients.
• Replacement of Batch Normalization (BN) with Filter Response Normalization (FRN) to eliminate batch size dependency.
• Development of a 3D extension, IResUnet3+, based on the 2D improved model for accurate 3D brain tumor segmentation.

Main Results:

• The improved network demonstrated increased sensitivity for Whole Tumor (WT), Tumor Core (TC), and Enhancing Tumor (ET) by 1.34%, 4.6%, and 8.44%, respectively.
• Dice coefficients for ET and WT were improved by 3.43% and 1.03%, respectively.
• The IResUnet3+ model achieved accurate segmentation of 3D brain tumor data.

Conclusions:

• The proposed IResUnet3+ network significantly improves brain tumor segmentation accuracy on the BraTS2018 dataset compared to U-Net, V-Net, ResUNet, and U-Net3+.
• The enhanced network features smaller parameters and superior accuracy, making it a promising tool for clinical applications.