This study introduces a knee cartilage segmentation method based on semi-supervised learning and conditional probability, to address the scarcity and quality issues of annotated samples in medical image segmentation. As it is difficult for existing embedded deep learning models to effectively model the hierarchical relationships among network outputs, the study proposes an approach combining conditional-to-unconditional mixed training and task-level consistency. In this way, the hierarchical relationships and relevance among labels are efficiently utilized, and the segmentation accuracy is enhanced. Specifically, the study employs a dual-task deep network predicting both pixel-level segmentation images and geometric perception level set representations of the target. The level set is shifted into an approximate segmentation map through a differentiable task transformation layer. Meanwhile, the study also introduces task-level consistency regularization between level line-based and directly predicted segmentation maps on labeled and unlabeled data. Extensive experiments on two public datasets demonstrate that this approach can significantly improve performance through the incorporation of unlabeled data.