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

Annotations of 3D medical images for segmentation require specialized expertise and are time-consuming, making large, labeled datasets rare and challenging to produce. Our objective was to investigate whether large unlabeled human datasets can be leveraged using cross-species self-supervised transfer learning to enhance the segmentation of pulmonary lobes in computed tomography (CT) scans from nonhuman primates with and without lower respiratory infection. A total of 1667 unlabeled human chest CT scans were assembled from two publicly available sources, and 23 chest CT scans of crab-eating macaques were annotated for the locations of the pulmonary lobes. The unlabeled human scans were used to train a 3D vision transformer (ViT) autoencoder in a self-supervised manner using contrastive learning. The pretrained ViT encoder was transferred to a U-Net transformers (UNETR) segmentation model, which was then trained using the labeled macaque dataset to perform pulmonary lobe segmentation. Ablation experiments on the effects of self-supervised pretraining, layer freezing, and data augmentation were conducted. The segmentation model with cross-species self-supervised pretraining achieved high performance (Dice similarity coefficient (DSC) = 90.31 ± 1.77) that was significantly greater than without pretraining (ΔDSC = 1.2%, t_paired = 5.3, p = 1.8E-3). Ablation experiments on the human pretraining data demonstrated that the amount of data and diversity of sources were important to performance. In fine-tuning, freezing just the first three layers of the ViT produced the best-performing model, and data augmentation before self-supervised pretraining and supervised fine-tuning was critical for high performance. Cross-species self-supervised transfer learning significantly improved the macaque pulmonary lobe segmentation performance with no additional acquisition or annotation costs.