Dilated multi-scale residual attention (DMRA) U-Net: three-dimensional (3D) dilated multi-scale residual attention U-Net for brain tumor segmentation

Lihong Zhang ¹, Yuzhuo Li ¹, Yingbo Liang ¹

⁰College of Computer Science and Technology, Henan Polytechnic University, Jiaozuo, China.

Quantitative Imaging in Medicine and Surgery +

|

October 21, 2024

View abstract on PubMed

Summary

This study introduces a novel 3D dilated multi-scale residual attention U-Net (DMRA-U-Net) for improved brain tumor segmentation. The DMRA-U-Net model enhances accuracy in identifying whole tumor, tumor core, and enhancing tumor regions in MRI scans.

Area Of Science

Medical Imaging
Artificial Intelligence
Neuro-oncology

Background

Accurate identification of brain tumor position and form is crucial for diagnosis and treatment.
Segregating brain gliomas and their sub-regions is challenging due to complex categories, sizes, and forms.
Existing methods struggle with precise brain tumor segmentation tasks.

Purpose Of The Study

To design a novel deep-learning network to overcome limitations in brain tumor segmentation.
To improve the accuracy of segmenting brain gliomas and their internal sub-regions.
To develop a 3D U-Net based model for enhanced brain tumor segmentation (BraTS).

Main Methods

Developed a 3D dilated multi-scale residual attention U-Net (DMRA-U-Net) model.
Incorporated dilated convolution residual (DCR) modules for shallow feature processing.
Utilized multi-scale convolution residual (MCR) modules and channel attention (CA) for comprehensive feature expression and retention.

Main Results

The DMRA-U-Net model achieved high segmentation performance on BraTS 2018-2021 datasets.
Achieved Dice Similarity Coefficients (DSC) of 0.9012 (WT), 0.8867 (TC), and 0.8813 (ET).
Demonstrated significant improvements over traditional 3D U-Net, with increased DSC and sensitivity, and decreased Hausdorff Distance.

Conclusions

The developed DMRA-U-Net is a promising model for brain tumor segmentation.
The model shows strong performance in segmenting whole tumor, tumor core, and enhancing tumor regions.
The solution is open-sourced and available for further research and application.

Keywords:

Brain tumor segmentation (BraTS) U-Net multi-scale convolution residual modules (MCR modules) three-dimensional (3D)