Related Experiment Video
Updated: Mar 22, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.7K
Decoupling Target Semantics via Text-Anchored Visual Contrast for Semi-Supervised Medical Image Segmentation.
Summary
This study introduces Text-anchored Visual Decoupling (TeViD), a new framework for semi-supervised medical image segmentation. TeViD improves accuracy by disentangling target and background information using both visual and textual data.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Semi-supervised learning (SSL) reduces annotation needs but struggles with ambiguity under limited supervision.
- Existing text-enhanced SSL methods often prioritize feature fusion over crucial target semantics for segmentation.
- Medical image segmentation requires accurate differentiation of target structures from background, especially with scarce labeled data.
Purpose of the Study:
- To propose a novel Text-anchored Visual Decoupling (TeViD) framework for semi-supervised medical image segmentation.
- To address semantic ambiguity and improve segmentation performance by leveraging both visual and textual information.
- To enhance the disentanglement of target and background representations in medical images.
Main Methods:
- A teacher-student architecture with a dual-decoder design to disentangle target and background representations.
- A reversed cross-supervision mechanism for unlabeled data to improve decoder diversity and semantic separation.
- Two contrastive learning objectives: teacher-guided visual contrastive loss and text-anchored contrastive loss for semantic reinforcement.
Main Results:
- TeViD consistently outperformed standard SSL and text-enhanced SSL methods across five diverse medical imaging datasets (X-ray, pathology, ultrasound, MRI, CT).
- Achieved average improvements of 5.72% in Dice score and 8.15% in mean Intersection over Union (mIoU) compared to the second-best method.
- Demonstrated effective semantic disentanglement from both visual and textual perspectives, enhancing segmentation accuracy.
Conclusions:
- The proposed TeViD framework offers a significant advancement in semi-supervised medical image segmentation.
- Leveraging text-anchored visual decoupling effectively addresses semantic ambiguity and improves segmentation performance.
- TeViD provides a robust and versatile solution for various medical imaging modalities.
