Related Experiment Video
Updated: Oct 10, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.0K
DCACNet: Dual context aggregation and attention-guided cross deconvolution network for medical image segmentation.
Hongchun Lu1, Shengwei Tian2, Long Yu3
1School of Software, Xinjiang University, Urumqi, Xinjiang 830046, China; School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, Sichuan 610031, China.
Computer Methods and Programs in Biomedicine
|December 10, 2021
Summary
This study introduces DCACNet, a novel deep learning model that enhances medical image segmentation by preserving deep feature information and addressing feature map sparsity, leading to improved performance.
Area of Science:
- Biomedical Image Analysis
- Deep Learning
- Medical Image Processing
Background:
- Standard Convolutional Neural Networks (CNNs) struggle with spatial information loss during coding, hindering detailed feature restoration and causing performance degradation in medical image segmentation.
- The sparsity of feature maps in existing models presents a significant challenge for accurate segmentation of complex medical images.
Purpose of the Study:
- To develop a deep learning framework, DCACNet, that effectively preserves deep image features for improved medical image segmentation.
- To overcome the limitations of standard CNNs in capturing fine-grained details and addressing feature map sparsity in segmentation tasks.
Main Methods:
- A novel deep learning network, DCACNet, was constructed for enhanced medical image segmentation.
- A multiscale cross-fusion encoding network was employed for robust feature extraction.
- A dual context aggregation module was integrated to fuse multi-scale contextual information and capture intricate deep features.
- An attention-guided cross deconvolution decoding network was utilized to generate dense feature maps, mitigating information loss.
Main Results:
- DCACNet demonstrated superior segmentation performance compared to existing models on public datasets.
- On the CHAOS dataset (4-class), the mean Dice Similarity Coefficient (DSC) reached 91.03%.
- On the Herlev dataset (2-class), DCACNet achieved high accuracy (96.77%), precision (90.40%), sensitivity (94.20%), specificity (97.50%), and Dice score (97.69%).
Conclusions:
- DCACNet significantly improves medical image segmentation performance by effectively retaining deep feature information.
- The proposed model successfully addresses the sparsity problem in medical image segmentation, offering a more robust solution.
- DCACNet's architecture enables better capture of fine-grained details, leading to more accurate segmentation outcomes.
