Random Convolutions for Domain Generalization of Deep Learning-based Medical Image Segmentation Models.
Authors
Affiliations (3)
Affiliations (3)
- Institute for Diagnostic and Interventional Neuroradiology, Klinikum rechts der Isar, Technical University of Munich, Neuro-Kopf-Zentrum, Ismaninger Str 22, 81675 Munich, Germany.
- Institute for Artificial Intelligence and Informatics in Medicine, Technical University of Munich, Munich, Germany.
- Department of Radiation Oncology, Klinikum rechts der Isar, Technical University of Munich, Munich, Germany.
Abstract
Purpose To evaluate random convolutions as an augmentation strategy for improving domain generalization of deep learning-based segmentation models in medical imaging. Materials and Methods In this retrospective study, a random-convolution-based augmentation strategy was applied to abdominal organ segmentation (AbdomenCT-1k: 361 CT images; AMOS: 298 CT and 59 MRI scans) and brain tissue segmentation (IXI: 504 T1-weighted [T1w] images from Guy's and Hammersmith Hospitals, 146 paired T1w/T2-weighted [T2w] images from the Institute of Psychiatry). Performance was compared with baseline and state-of-the-art segmentation models (TotalSegmentator and deepAtropos). Random convolution configurations were analyzed for effects on in-and out-of-domain performance. Results The random convolution-enhanced UNet achieved in-domain Dice scores comparable to state-of-the-art baselines (CT: 0.93 vs TotalSegmentator: 0.95; T1w imaging: 0.83 vs deepAtropos: 0.79). Out-of-domain Dice scores were significantly higher (MRI: 0.93, T2w imaging: 0.52) compared with baselines (TotalSegmentator in MRI: 0.85, deepAtropos in T2w imaging: 0.33, FDR-adjusted <i>P</i> values < 0.001). Augmentation probability and configuration influenced the trade-off between in-and out-of-domain performance. Conclusion Random convolutions yielding more robust segmentation models that generalized better to unseen domains than models trained without random convolutions and are compatible with diverse segmentation architectures. ©RSNA, 2025.