Cross-Species Self-supervised Transfer Learning for Pulmonary Lobe Segmentation in Nonhuman Primates

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• 3D medical image segmentation, particularly for pulmonary lobes in CT scans, is hindered by the scarcity of expert-annotated datasets.
• Developing accurate segmentation models for nonhuman primates is crucial for comparative research and understanding respiratory diseases.

Purpose of the Study:

• To investigate the efficacy of cross-species self-supervised transfer learning using unlabeled human CT data to improve pulmonary lobe segmentation in macaque CT scans.
• To evaluate the impact of pretraining strategies, layer freezing, and data augmentation on segmentation performance.

Main Methods:

• A 3D vision transformer (ViT) autoencoder was trained using contrastive learning on 1667 unlabeled human chest CT scans.
• The pretrained ViT encoder was integrated into a U-Net transformers (UNETR) model and fine-tuned on 23 annotated macaque CT scans for pulmonary lobe segmentation.
• Ablation studies were performed to assess the contributions of self-supervised pretraining, data augmentation, and layer freezing during fine-tuning.

Main Results:

• The UNETR model with cross-species self-supervised pretraining achieved a high Dice Similarity Coefficient (DSC) of 90.31% ± 1.77%.
• This performance was significantly superior to models without pretraining (ΔDSC = 1.2%, p < 0.002).
• Data quantity and diversity for pretraining, freezing the first three ViT layers, and data augmentation were critical factors for optimal performance.

Conclusions:

• Cross-species self-supervised transfer learning effectively enhances pulmonary lobe segmentation in nonhuman primate CT scans.
• This method offers a cost-effective solution by leveraging readily available unlabeled human data, bypassing the need for extensive macaque-specific annotations.
• The findings highlight the potential of self-supervised learning for medical image analysis across different species.