Learnable weight initialization for volumetric medical image segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Hybrid models combining local convolution and global attention are popular for volumetric medical image segmentation.
• Current methods often use data-independent weight initialization, limiting performance by not leveraging inherent data characteristics.

Purpose of the Study:

• To propose a learnable, data-dependent weight initialization approach for hybrid volumetric medical image segmentation models.
• To improve segmentation performance by effectively learning contextual and structural cues from medical training data.

Main Methods:

• Developed a novel learnable weight initialization strategy using self-supervised objectives.
• Integrated the approach into existing hybrid models without requiring external datasets.
• Evaluated performance on multi-organ and lung cancer segmentation tasks.

Main Results:

• Achieved state-of-the-art segmentation performance on tested tasks.
• Demonstrated superior performance compared to Swin-UNETR pretrained on large datasets for multi-organ segmentation.
• The proposed method is easy to integrate and requires no external training data.

Conclusions:

• The proposed data-dependent weight initialization significantly enhances hybrid volumetric medical image segmentation.
• This approach effectively utilizes training data to capture crucial image cues, leading to improved accuracy.
• The method offers a practical and effective solution for advancing medical image analysis.