Retinal layer segmentation in optical coherence tomography (OCT) images helps to quantify and diagnose many ocular diseases. However, the performance of existing methods is mostly constrained by limited dense annotations. Additionally, the layer deformation and blurred boundaries in OCT images make layer segmentation even more challenging. To alleviate these issues, a Semi-supervised layer segmentation method based on Cross-task Consistency and Contrastive Learning (SC3L-Net) is proposed. Specifically, a wavelet-guided multi-scale vertical attention module that can perceive global information is proposed to alleviate the layer deformation problem caused by lesions. Then, the method introduces a regression task based on the segmentation task and constructs cross-task consistency learning, aiming to alleviate the boundary deviation caused by blurred boundaries. Finally, to further enhance the model’s segmentation outcomes, a pixel-wise contrastive loss is incorporated to improve the feature representation of unlabeled data. The proposed SC3L-Net achieved strong performance, surpassing the comparison methods on the OCT5K-AMD dataset.