Related Experiment Video
Updated: Jul 8, 2026
07:35
Selecting Multiple Biomarker Subsets with Similarly Effective Binary Classification Performances
Published on: October 11, 2018
DreamOn: a data augmentation strategy to narrow the robustness gap between expert radiologists and deep learning
Luc Lerch1,2, Lukas S Huber3,4, Amith Kamath1
1Medical Image Analysis Group, ARTORG Centre for Biomedical Research, University of Bern, Bern, Switzerland.
Frontiers in Radiology
|January 6, 2025
Summary
Deep learning models for medical imaging require robust performance against image noise. A novel data augmentation strategy, DreamOn, inspired by REM dreams, significantly improves AI robustness in noisy conditions, though human radiologists still outperform AI.
Area of Science:
- Medical Image Analysis
- Artificial Intelligence
- Deep Learning
Background:
- Deep learning model performance in medical imaging is contingent on image quality.
- Variability in imaging equipment, calibration, and patient factors introduces noise, impacting diagnostic accuracy.
- Robustness against noise is critical for clinical deployment of AI in healthcare.
Purpose of the Study:
- To assess the impact of data augmentation on the noise robustness of a ResNet-18 model for breast ultrasound image classification.
- To compare the performance of augmented deep learning models against human radiologists.
- To introduce and evaluate DreamOn, a novel biologically inspired data augmentation technique.
Main Methods:
- ResNet-18 model trained on breast ultrasound images.
- Evaluation of various data augmentation strategies for noise robustness.
- Benchmarking AI performance against trained human radiologists.
- Implementation of DreamOn, a conditional Generative Adversarial Network (GAN) for generating REM-dream-inspired image interpolations.
Main Results:
- Standard data augmentation methods enhance model robustness compared to no augmentation.
- Human radiologists demonstrate superior performance on noisy images compared to AI models.
- The DreamOn data augmentation strategy significantly improves model robustness, particularly in high-noise scenarios.
Conclusions:
- REM-dream-inspired, GAN-based data augmentation (DreamOn) shows promise for enhancing AI robustness against noise in medical imaging.
- A notable performance gap persists between current AI models and human experts in handling noisy medical images.
- Further advancements in AI are necessary to achieve human-level diagnostic expertise and robustness in clinical applications.
