Related Experiment Video
Updated: Aug 20, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
E-DU: Deep neural network for multimodal medical image segmentation based on semantic gap compensation.
Haojia Wang1, Xicheng Chen1, Rui Yu2
1Department of Health Statistics, College of Preventive Medicine, Army Medical University, NO.30 Gaotanyan Street, Shapingba District, Chongqing, 400038, China.
Computers in Biology and Medicine
|November 17, 2022
Summary
This study introduces an enhanced dense U-Net (E-DU) with a novel multiscale denoise enhancement (MDE) module to improve multimodal medical image segmentation. The E-DU model effectively addresses semantic gaps, enhancing segmentation accuracy and efficiency across various medical imaging modalities.
Area of Science:
- Medical image analysis
- Deep learning for medical imaging
- Computer-aided diagnosis
Background:
- U-Net is a benchmark for medical image segmentation but struggles with semantic gaps in traditional skip connections.
- Direct feature fusion can lead to fuzzy feature maps and segmentation errors.
Purpose of the Study:
- To improve multimodal medical image segmentation performance and efficiency.
- To address the semantic gap issue in U-Net architectures.
Main Methods:
- Propose an enhanced dense U-Net (E-DU) incorporating a multiscale denoise enhancement (MDE) module.
- Replace traditional skip connections with the MDE module for spatial enhancement filtering.
- Develop a deep full convolution network structure for feature fusion, denoising, and enhancement.
Main Results:
- E-DU achieved superior segmentation results across seven medical image modalities, outperforming the U-Net family.
- Demonstrated significant improvements in Dice Similarity Coefficient (DSC) values, reaching up to 98.85%.
- The MDE module enhanced segmentation performance and efficiency when added to attention mechanism networks.
Conclusions:
- The proposed MDE module offers effective and efficient segmentation for multimodal medical images.
- The E-DU model shows great potential for clinical applications, aiding decision support through comprehensive image information utilization.
- The method is easily extensible to diverse medical segmentation datasets and advanced architectures.
