Domain and Content Adaptive Convolution Based Multi-Source Domain Generalization for Medical Image Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Variable medical image quality creates a domain gap, hindering the clinical application of segmentation models trained in labs.
• Existing domain generalization methods often rely on static convolutions, limiting their flexibility and adaptability.

Purpose of the Study:

• To propose a multi-source domain generalization model, Domain and Content Adaptive Convolution (DCAC), for robust medical image segmentation across diverse modalities.
• To enhance model adaptability to unseen clinical data and varying image characteristics.

Main Methods:

• Developed Domain Adaptive Convolution (DAC) and Content Adaptive Convolution (CAC) modules integrated into an encoder-decoder architecture.
• DAC uses a dynamic convolutional head conditioned on predicted domain codes for target domain adaptation.
• CAC employs a dynamic convolutional head conditioned on global image features for test image adaptation.

Main Results:

• The DCAC model significantly outperformed baseline and state-of-the-art domain generalization methods on prostate, COVID-19 lesion, and optic cup/disc segmentation tasks.
• Demonstrated the individual and combined effectiveness of the DAC and CAC modules in improving segmentation performance.
• Achieved superior results across multiple segmentation tasks, highlighting the model's generalizability.

Conclusions:

• The proposed DCAC model effectively addresses the domain gap in medical image segmentation.
• The adaptive convolution modules (DAC and CAC) are crucial for enhancing model performance and adaptability.
• DCAC offers a flexible and powerful solution for deploying segmentation models in diverse clinical settings.