Related Experiment Video
Updated: Jun 4, 2026
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
Frequency disentanglement with State space gating network for medical image segmentation
Zuo Huang1,2, Xiang Li1,2, Jinyu Cong1,2
1Center for Medical Artificial Intelligence, Shandong University of Traditional Chinese Medicine, Qingdao Traditional Chinese Medicine Inheritance and Innovation Base, East Side of Fenghe Road, Qingdao High-tech Industrial Development Zone, Qingdao, 266112, Shandong, China.
This study introduces FD-SSGNet, a novel framework for medical image segmentation that disentangles spectral features using Fast Fourier Transform (FFT) and State-Space gating, significantly improving accuracy.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Artificial Intelligence
Background:
- Automated segmentation of anatomical structures is crucial for clinical applications like computer-aided diagnosis and radiotherapy planning.
- Current models (CNNs, Transformers) struggle with spectral feature entanglement, mixing global, contour, and texture information, which degrades segmentation accuracy, especially at object boundaries.
- This limitation hinders precise delineation required for quantitative medical analysis.
Purpose of the Study:
- To develop a novel framework, FD-SSGNet, that explicitly addresses spectral feature entanglement in medical image segmentation.
- To improve the accuracy and robustness of automated segmentation by disentangling frequency components.
- To enhance clinical workflows through more precise anatomical structure delineation.
Main Methods:
- The FD-SSGNet framework utilizes the Fast Fourier Transform (FFT) to decompose feature maps into low, mid, and high-frequency components.
- It employs a Shift Bidirectional Selective Gate Mamba (SBSGM) with parallel pathways to model frequency-specific long-range dependencies.
- A dynamic fusion module adaptively reintegrates these processed multi-band features for refined segmentation.
Main Results:
- FD-SSGNet achieved new state-of-the-art performance on the BTCV multi-organ and ACDC cardiac segmentation datasets.
- Explicit modeling in the frequency domain significantly improved segmentation accuracy, particularly at critical object boundaries.
- The framework demonstrated robust and accurate medical image analysis capabilities.
Conclusions:
- Explicit frequency domain modeling offers significant benefits for robust and accurate medical image segmentation.
- FD-SSGNet represents a advancement in automated medical image analysis, enhancing clinical workflows.
- The proposed method validates the importance of addressing spectral feature entanglement for precise segmentation.
