Related Experiment Video
Updated: Mar 21, 2026
04:48
Swin-PSAxialNet: An Efficient Multi-Organ Segmentation Technique
Published on: July 5, 2024
854
Wavelet-Transformed Frequency Linked Attention with Selective Hierarchy for Abdominal Organ Segmentation
Chung-Ming Lo1, Kai-Chun Huang2, Chuan-Hsien Chen2
1Institute of Artificial Intelligence Innovation, Industry Academia Innovation School, National Yang Ming Chiao Tung University, Hsinchu, Taiwan.
Journal of Imaging Informatics in Medicine
|March 20, 2026
Summary
A new deep learning model called Frequency Linked Attention with Selective Hierarchy (FLASH) improves automatic organ segmentation in abdominal CT scans. This method enhances accuracy and efficiency for better clinical applications.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer Vision
Background:
- Computed tomography (CT) scans are increasingly used for detailed 3D organ visualization and disease diagnosis.
- Accurate quantitative evaluation and automatic segmentation of CT scans are crucial for effective clinical decision-making but remain challenging.
- Existing segmentation methods struggle with unclear organ boundaries in abdominal CT scans due to low contrast.
Purpose of the Study:
- To introduce a novel deep learning architecture, Frequency Linked Attention with Selective Hierarchy (FLASH), for precise multi-organ segmentation in abdominal CT scans.
- To address the limitations of current segmentation techniques, particularly in handling low-contrast images and unclear boundaries.
- To evaluate the performance of FLASH against other networks in terms of accuracy, efficiency, and resource utilization.
Main Methods:
- Developed FLASH, a deep learning model incorporating 3D discrete wavelet transform within Transformer blocks for enhanced feature extraction.
- Utilized frequency band distinction to overcome low-contrast challenges and improve organ boundary definition.
- Implemented two-directional skip connections and attention mechanisms to integrate multi-scale frequency features and adaptively weight components.
Main Results:
- FLASH achieved superior segmentation performance, evidenced by the highest Dice Similarity Coefficient (0.826) and lowest Normalized Surface Distance (0.698).
- The model demonstrated excellent consistency with low standard deviations for both metrics (DSC: 0.034, NSD: 0.046).
- Compared to Swin UNETR, FLASH exhibited fewer parameters (96.86M vs. 139.45M) and reduced computation time (36.3h vs. 50.4h).
Conclusions:
- FLASH offers a significant advancement in automatic multi-organ segmentation for abdominal CT scans.
- The model's improved accuracy and efficiency make it highly suitable for clinical settings with resource and time constraints.
- The open-source availability of the FLASH code facilitates further research and clinical adoption.
