Related Experiment Video
Updated: Jun 30, 2025
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
PFD-Net: Pyramid Fourier Deformable Network for medical image segmentation
Chaorong Yang1, Zhaohui Zhang1
1College of Computer and Cyber Security, Hebei Normal University, Shijiazhuang 050024, China; Hebei Provincial Engineering Research Center for Supply Chain Big Data Analytics & Data Security, Shijiazhuang 050024, China; Hebei Provincial Key Laboratory of Network & Information Security, Hebei Normal University, Shijiazhuang 050024, China.
The novel Pyramid Fourier Deformable Network (PFD-Net) improves medical image segmentation by integrating local and global details, outperforming existing methods on diverse datasets for tasks like polyp and cardiac segmentation.
Area of Science:
- Computer Vision
- Medical Imaging Analysis
- Deep Learning Architectures
Background:
- Accurate medical image segmentation is vital for clinical diagnosis, yet current methods struggle to integrate local details and global semantic information.
- This limitation hinders the effective segmentation of fine-grained content, including small targets and irregular boundaries.
Purpose of the Study:
- To introduce the Pyramid Fourier Deformable Network (PFD-Net), a novel deep learning model for enhanced medical image segmentation.
- To address the limitations of existing methods in capturing both local details and global semantic context.
Main Methods:
- The PFD-Net employs a PVTv2-based Transformer encoder for global information capture, enhanced by the Fast Fourier Convolution Residual (FFCR) module for local and global feature representation.
- The Dilated Deformable Refinement (DDR) module is introduced to improve comprehension of global semantic structures for diverse targets and irregular boundaries.
- A Cross-Level Fusion Block (CLFB) with deformable convolution integrates features from the decoder and an auxiliary CNN encoder to enhance target perception.
Main Results:
- PFD-Net demonstrated superior performance across nine medical image datasets and five segmentation tasks (polyp, abdominal, cardiac, gland cells, nuclei).
- Achieved a leading mDice score of 0.826 on the challenging ETIS dataset, surpassing previous state-of-the-art methods.
- Showcased significant improvements, including 1.8% and 3.6% gains over HSNet and PVT-CASCADE, respectively.
Conclusions:
- PFD-Net effectively leverages both CNN and Transformer strengths for superior medical image segmentation.
- The proposed architecture significantly enhances the ability to segment complex structures and irregular boundaries in medical images.
- PFD-Net represents a substantial advancement in medical image segmentation accuracy and robustness.
