Towards liver segmentation in the wild via contrastive distillation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Surgery

Background:

• Automatic liver segmentation is crucial for computer-assisted hepatic procedures.
• High variability in organ appearance and limited labels pose significant challenges.
• Existing supervised methods struggle with generalization to unseen data.

Purpose of the Study:

• To develop a robust automatic liver segmentation method with strong generalization capabilities.
• To overcome the limitations of existing supervised methods in real-world scenarios.
• To enable accurate liver segmentation across diverse imaging modalities and patient data.

Main Methods:

• A novel contrastive distillation scheme is proposed, leveraging a pre-trained large neural network.
• Neighboring slices are mapped closely in latent representation, while distant slices are mapped farther apart.
• A U-Net style upsampling path is trained using ground-truth labels to recover segmentation maps.

Main Results:

• The method demonstrates state-of-the-art inference performance on unseen domains.
• Extensive validation across six abdominal datasets and 18 patient datasets confirms robustness.
• Sub-second inference time and data-efficient training facilitate real-world scalability.

Conclusions:

• A novel contrastive distillation scheme for automatic liver segmentation is presented.
• The proposed method exhibits superior performance compared to state-of-the-art techniques.
• Its limited assumptions and robust performance make it suitable for real-world clinical applications.