Weakly-Supervised Self-Ensembling Vision Transformer for MRI Cardiac Segmentation

Deep learning techniques are crucial in medical image segmentation, but their effectiveness heavily relies on a vast amount of fully annotated data, which is costly in labour and time. To address this challenge, this paper introduced a framework for scribble-supervised learning with a self-ensembling approach. The transformation consistency scheme is further developed to boost performance. Inspired by the recent achievements of the Vision Transformer (ViT) in modeling long-range dependencies, and to enable a fair comparison with convolutional operations, we employ a U-shaped segmentation network composed of pure self-attention-based blocks. Our proposed scribble-supervised segmentation ViT is validated on a public benchmark dataset against classical methods with various evaluation metrics. The code, trained model, and preprocessed scribble-annotated sets are publicly available at https://github.com/ziyangwang007/CV-WSL-MIS.