Learning generalizable visual representation via adaptive spectral random convolution for medical image segmentation

Zuyu Zhang, Yan Li, Byeong-Seok Shin

Abstract: Highlights•Proposes an innovative ASRConv framework, leveraging noise-adaptive random convolution to enhance feature variability.•Features a unique weight generation module with a noise-adaptive head, ensuring robust feature extraction amidst varying noise levels.•Implements an adversarial domain augmentation strategy for adaptive suppression of high-frequency noises.•Demonstrates superior performance over existing methods, indicating robust generalization capability across unseen target domains.