Fig. 6: Detailed architecture of CereVessPro and CereVessSeg for cerebrovascular segmentation.

The encoder and the first two stages of the decoder from the segmentation model CereVessSeg are initially pretrained using the contrastive learning method CereVessPro with a large-scale unannotated dataset. Afterward, the entire CereVessSeg model is finetuned using the annotated dataset.