DreamOn: a data augmentation strategy to narrow the robustness gap between expert radiologists and deep learning classifiers
Luc Lerch, Lukas S. Huber, Amith J. Kamath, Alexander Pöllinger, Aurélie Pahud de Mortanges, Verena Carola Obmann, Florian Dammann, Walter Senn, Mauricio Reyes · Frontiers in Radiology · 2024
Purpose: Successful performance of deep learning models for medical image analysis is highly dependent on the quality of the images being analysed. Factors like differences in imaging equipment and calibration, as well as patient-specific factors such as movements or biological variability (e.g., tissue density), lead to a large variability in the quality of obtained medical images. Consequently, robustness against the presence of noise is a crucial factor for the application of deep learning models in clinical contexts. Materials and methods: , a novel, biologically inspired data augmentation strategy for medical image analysis. DreamOn is based on a conditional generative adversarial network (GAN) to generate REM-dream-inspired interpolations of training images. Results: We find that while available data augmentation approaches substantially improve robustness compared to models trained without any data augmentation, radiologists outperform models on noisy images. Using DreamOn data augmentation, we obtain a substantial improvement in robustness in the high noise regime. Conclusions: We show that REM-dream-inspired conditional GAN-based data augmentation is a promising approach to improving deep learning model robustness against noise perturbations in medical imaging. Additionally, we highlight a gap in robustness between deep learning models and human experts, emphasizing the imperative for ongoing developments in AI to match human diagnostic expertise.