Related Experiment Video
Updated: Jul 1, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
PAC-Net: patch adaptive cut-off network with differentiable module-wise K-learning for robust and efficient medical
Xiang Pan1, Weiming Zhu1, Herong Zheng1
1ZheJiang University of Technology, Hangzhou, People's Republic of China.
Physics in Medicine and Biology
|June 29, 2026
Summary
This study introduces PACNet, a novel medical image segmentation network with an adaptive attention mechanism. PACNet improves segmentation accuracy and robustness by learning data-dependent sparsity, outperforming existing methods.
Area of Science:
- Medical Image Analysis
- Computer Vision
- Machine Learning
Background:
- Sparse attention networks in medical image segmentation often use fixed K values, limiting adaptation to diverse lesion sizes and shapes.
- This rigidity hinders robustness and segmentation accuracy in medical image analysis systems.
Purpose of the Study:
- To develop an adaptive, differentiable sparse attention mechanism for medical image segmentation.
- To enhance the robustness and segmentation accuracy of medical image analysis systems.
Main Methods:
- Proposed PACNet (Patch Adaptive Cut-off Network) with an Entropy-Guided Differentiable K-Selection (EGDK) module.
- EGDK learns data-dependent sparsity ratios using Gaussian Soft Indexing and a Straight-Through Estimator (STE) for end-to-end differentiability.
- Evaluated on eight datasets across five imaging modalities.
Main Results:
- PACNet achieved a superior average Dice Similarity Coefficient (DSC) of 90.57% across eight datasets.
- Outperformed BRAU-Net++ by +1.79% in average DSC.
- Demonstrated efficiency with 6.82M parameters and 6.02G FLOPs.
Conclusions:
- Adaptive and differentiable K selection is superior to fixed or discrete methods for medical image segmentation.
- PACNet effectively reduces background noise and preserves fine anatomical details.
- Presents a practical and clinically relevant solution for robust medical image segmentation.