Random Convolutions for Domain Generalization of Deep Learning-based Medical Image Segmentation Models

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning models for medical image segmentation often struggle with domain generalization.
• Augmentation strategies are crucial for improving model robustness and performance on unseen data.

Purpose of the Study:

• To evaluate random convolutions as an augmentation technique for enhancing domain generalization in deep learning-based medical image segmentation.
• To assess the impact of random convolutions on segmentation performance across different imaging modalities and datasets.

Main Methods:

• A retrospective study applying a random-convolution-based augmentation strategy to abdominal organ and brain tissue segmentation tasks.
• Performance comparison of the augmented UNet model against baseline and state-of-the-art segmentation models (TotalSegmentator, deepAtropos).
• Analysis of random convolution configurations for their effect on in-domain and out-of-domain performance.

Main Results:

• The random convolution-enhanced UNet achieved competitive in-domain Dice scores comparable to state-of-the-art models.
• Significantly higher out-of-domain Dice scores were observed for the augmented model on MRI and T2w imaging compared to baselines (FDR-adjusted P < 0.001).
• Augmentation probability and configuration influenced the balance between in-domain and out-of-domain performance.

Conclusions:

• Random convolutions yield more robust medical image segmentation models with improved generalization to unseen domains.
• This augmentation strategy is compatible with diverse deep learning segmentation architectures.
• The findings suggest random convolutions as a valuable tool for enhancing the reliability of AI in medical imaging.