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

Medical image segmentation models often fail to generalize well when applied to new datasets, hindering their usage in clinical practice. Existing random-convolution-based domain generalization approaches, which involve randomizing the convolutional kernel weights in the initial layers of CNN models, have shown promise in improving model generalizability. Nevertheless, the indiscriminate introduction of high-frequency noise during early feature extraction may pollute the critical fine details and degrade the model's performance on new datasets. To mitigate this problem, we propose an adaptive spectral random convolution (ASRConv) module designed to selectively randomize low-frequency features while avoiding the introduction of high-frequency artifacts. Unlike prior arts, ASRConv dynamically generates convolution kernel weights, enabling more effective control over feature frequencies than randomized kernels. Specifically, ASRConv achieves this selective randomization through a novel weight generation module conditioned on random noise inputs. The adversarial domain augmentation strategy guides the weight generation module in adaptively suppressing high-frequency noise during training, allowing ASRConv to improve feature diversity and reduce overfitting to specific domains. Extensive experimental results show that our proposed ASRConv method consistently outperforms the state-of-the-art methods, with average DSC improvements of 3.07% and 1.18% on fundus and polyp datasets, respectively. We also qualitatively demonstrate the robustness of our model against domain distribution shifts. All these results validate the effectiveness of the proposed ASRConv in learning domain-invariant representations for robust medical image segmentation.