Related Experiment Video
Updated: Jan 7, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Semi-supervised medical image classification via feature-level multi-scale consistency and adversarial training.
Li Shiyan1, Wang Shuqin1, Gu Xin1
1School of Computer and Information Engineering, Tianjin Normal University, No. 393, Binshui West Road, Xiqing District, Tianjin, 300387, Xiqing District, China.
Summary
This study introduces Feature-level multi-scale Consistency and Adversarial Training (FCAT), a novel semi-supervised learning framework for medical image analysis. FCAT enhances feature utilization and training stability, improving performance across diverse datasets.
Area of Science:
- Medical Image Analysis
- Machine Learning
- Computer Vision
Background:
- Semi-supervised learning (SSL) shows promise in medical imaging with limited annotations.
- Existing SSL methods face limitations in feature utilization and training stability.
Purpose of the Study:
- To propose a novel semi-supervised framework, Feature-level multi-scale Consistency and Adversarial Training (FCAT), to address limitations in current methods.
- To improve feature-level information utilization and training stability in medical image analysis.
Main Methods:
- Introduced a multi-scale feature-level consistency mechanism for robust feature alignment without external sample pools.
- Designed a bidirectional feature perturbation strategy within a teacher-student scheme for mutual consistency.
- Developed an intrinsic evaluation mechanism using entropy and complementary confidence to prioritize informative samples.
Main Results:
- FCAT demonstrated competitive performance on balanced and imbalanced datasets, including Pneumonia Chest X-ray, NCT-CRC-HE, and ISIC 2019.
- Achieved strong generalization across diverse medical imaging modalities.
- Showcased improved feature utilization and training stability compared to existing methods.
Conclusions:
- FCAT offers a robust and stable semi-supervised learning approach for medical image analysis.
- The proposed methods effectively leverage feature-level information and handle data heterogeneity.
- FCAT presents a significant advancement for SSL in medical imaging applications.