Related Experiment Video
Updated: Jul 2, 2025
10:44
Three-Dimensional Phase Resolved Functional Lung Magnetic Resonance Imaging
Published on: June 21, 2024
498
Approximating Intermediate Feature Maps of Self-Supervised Convolution Neural Network to Learn Hard Positive
Kyungjin Cho1, Ki Duk Kim2, Jiheon Jeong1
1Department of Bioengineering, Asan Medical Institute of Convergence Science and Technology, Asan Medical Center, 88 Olympic-Ro 43-Gil Songpa-Gu, Seoul, 05505, South Korea.
Journal of Imaging Informatics in Medicine
|February 21, 2024
Summary
Intermediate Feature Approximation (IFA) loss enhances contrastive learning for medical images by improving positive representations without extra augmentations. This method boosts performance in tasks like classification and GAN inversion, especially with limited data.
Area of Science:
- Artificial Intelligence
- Medical Imaging Analysis
- Deep Learning
Background:
- Contrastive learning advances deep learning but faces challenges with positive representations in medical imaging.
- Strong augmentations can disrupt contrastive learning for subtle medical image differences.
Purpose of the Study:
- To introduce Intermediate Feature Approximation (IFA) loss for improved contrastive learning in medical imaging.
- To enhance positive representation focus without additional data augmentation.
Main Methods:
- Proposed IFA loss to maximize cosine similarity between intermediate feature maps of original data and positive pairs.
- Combined IFA loss with InfoNCE loss for comprehensive representation learning.
- Evaluated performance on classification, object detection, and Generative Adversarial Network (GAN) inversion tasks.
Main Results:
- IFA loss significantly improved contrastive convolutional neural network performance.
- Demonstrated effectiveness in overcoming data imbalance and scarcity.
- Showcased utility as a perceptual loss encoder for GAN inversion.
Conclusions:
- IFA loss offers a novel approach to enhance contrastive learning for medical images.
- The method improves model robustness and performance across diverse downstream tasks.
- Publicly released model encourages further research and collaboration.
