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

Area of Science:

• Medical image analysis
• Computer vision
• Machine learning

Background:

• Medical image segmentation models struggle with generalizability across different datasets, limiting clinical applications.
• Current random-convolution methods improve generalization but can introduce high-frequency noise, degrading performance.
• Existing approaches lack precise control over feature frequency randomization.

Purpose of the Study:

• To develop a novel module, Adaptive Spectral Random Convolution (ASRConv), for improved domain generalization in medical image segmentation.
• To mitigate the negative impact of high-frequency noise introduced by traditional random convolution methods.
• To enhance the robustness and clinical applicability of deep learning models for medical image segmentation.

Main Methods:

• Proposed an Adaptive Spectral Random Convolution (ASRConv) module that selectively randomizes low-frequency features.
• Introduced a dynamic weight generation module conditioned on random noise for adaptive kernel generation.
• Employed an adversarial domain augmentation strategy to guide noise suppression and improve feature diversity.

Main Results:

• ASRConv achieved average Dice Similarity Coefficient (DSC) improvements of 3.07% on fundus datasets and 1.18% on polyp datasets.
• The method demonstrated consistent outperformance compared to state-of-the-art domain generalization techniques.
• Qualitative results confirmed the model's robustness against domain distribution shifts.

Conclusions:

• The proposed ASRConv module effectively learns domain-invariant representations for robust medical image segmentation.
• ASRConv offers superior control over feature frequencies, avoiding detrimental high-frequency artifacts.
• This approach significantly enhances model generalizability and performance in clinical settings.