CADS: A Self-Supervised Learner via Cross-Modal Alignment and Deep Self-Distillation for CT Volume Segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Self-supervised learning (SSL) advances annotation-efficient learning but has limitations in CT volume segmentation.
• Existing SSL methods rarely utilize multi-modal information and offer limited supervision depth.
• These limitations hinder the encoder's ability to effectively characterize CT volumes.

Purpose of the Study:

• To propose a novel self-supervised learning framework, CADS, addressing limitations in CT volume segmentation.
• To improve the encoder's representation learning by incorporating cross-modal alignment and deep self-distillation.
• To enhance the performance of 3D CT volume segmentation models.

Main Methods:

• Developed a pretext task for cross-modal alignment between 3D CT volumes and 2D X-ray images.
• Extended self-distillation to deep self-distillation, providing supervision to multiple encoder layers.
• Constructed a PVT-UNet model using the CADS-pretrained encoder for downstream segmentation tasks.

Main Results:

• CADS demonstrated lower computational complexity and GPU memory usage during pre-training compared to other SSL methods.
• The PVT-UNet model, pre-trained with CADS, achieved superior performance on seven downstream 3D CT volume segmentation tasks.
• Outperformed state-of-the-art SSL methods (MOCOv3, DiRA) and medical image segmentation methods (nnUNet, CoTr).

Conclusions:

• CADS effectively improves encoder characterization for 3D CT volumes through cross-modal alignment and deep self-distillation.
• The proposed method offers a more efficient and effective approach to self-supervised learning in medical image segmentation.
• CADS-pretrained models show significant potential for advancing 3D CT volume segmentation accuracy and efficiency.