Related Experiment Video
Updated: Jun 28, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
GMAlignNet: multi-scale lightweight brain tumor image segmentation with enhanced semantic information consistency
Jianli Song1, Xiaoqi Lu1,2, Yu Gu1
1Inner Mongolia Key Laboratory of Pattern Recognition and Intelligent Image Processing, School of Digital and Intelligent Industry, Inner Mongolia University of Science and Technology, Baotou 014010, People's Republic of China.
GMAlignNet improves brain tumor segmentation by using Ghost convolutions and a feature alignment unit to capture multi-scale information and correct misalignments. This lightweight model achieves high accuracy on the BraTS dataset, enhancing edge detail recognition.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- U-Net architectures are standard for brain tumor segmentation but suffer from information loss and feature misalignment due to repeated convolutions and sampling.
- Traditional convolutional blocks struggle to capture multi-scale information essential for accurate brain tumor segmentation.
Purpose of the Study:
- To propose GMAlignNet, a novel multi-scale feature-aligned segmentation model to address information loss and feature misalignment in brain tumor segmentation.
- To enhance the capture of multi-scale contextual information and improve segmentation accuracy using efficient operations.
Main Methods:
- Employed Ghost convolution, Ghost hierarchical decoupled fusion units, and Ghost hierarchical decoupled units in the encoding and decoding paths, replacing standard convolutions.
- Introduced a feature alignment unit to guide upsampled feature recovery using semantic information flow, correcting pixel-level misalignments.
- Optimized existing networks (DMFNet, HDCNet, 3D UNet) with the proposed method to validate its effectiveness.
Main Results:
- GMAlignNet achieved Dice coefficients of 81.65% (enhancing tumor), 90.07% (whole tumor), and 85.16% (tumor core) on the BraTS 2018 dataset.
- The model demonstrated computational efficiency with only 0.29 M parameters and 26.88G FLOPs, indicating a lightweight design.
- Experiments across BraTS 2018, 2019, and 2020 datasets showed improved handling of edge details and contour recognition.
Conclusions:
- GMAlignNet effectively addresses limitations of U-Net architectures in brain tumor segmentation by leveraging Ghost convolutions and feature alignment.
- The proposed model offers a computationally efficient and lightweight solution with superior performance in capturing multi-scale features and refining segmentation boundaries.
