Efficient few-shot medical image segmentation via self-supervised variational autoencoder

Few-shot medical image segmentation typically uses a joint model for registration and segmentation. The registration model aligns a labeled atlas with unlabeled images to form initial masks, which are then refined by the segmentation model. However, inevitable spatial misalignments during registration can lead to inaccuracies and diminished segmentation quality. To address this, we developed EFS-MedSeg, an end-to-end model using two labeled atlases and few unlabeled images, enhanced by data augmentation and self-supervised learning. Initially, EFS-MedSeg applies a 3D random regional switch strategy to augment atlases, thereby enhancing supervision in segmentation tasks. This not only introduces variability to the training data but also enhances the model’s ability to generalize and prevents overfitting, resulting in natural and smooth label boundaries. Following this, we use a variational autoencoder for a weighted reconstruction task, focusing the model’s attention on areas with lower Dice scores to ensure accurate segmentation that conforms to the atlas image’s shape and structural appearance. Moreover, we introduce a self-contrastive module aimed at improving feature extraction, guided by anatomical structure priors, thus enhancing the model’s convergence and segmentation accuracy. Results on multi-modal medical image datasets show that EFS-MedSeg achieves performance comparable to fully-supervised methods. Moreover, it consistently surpasses the second-best method in Dice score by 1.4%, 9.1%, and 1.1% on the OASIS, BCV, and BCH datasets, respectively, highlighting its robustness and adaptability across diverse datasets. The source code will be made publicly available at: https://github.com/NoviceFodder/EFS-MedSeg.