Abstract: Weakly supervised semantic segmentation based on scribble annotations faces two major challenges: pseudo-labels used for model training are highly susceptible to noise, and existing methods exhibit limited capability in recognizing irregular targets. These challenges are particularly evident in cardiac image segmentation tasks. To address these issues, a multi-layered, multi-channel semantic segmentation framework based on label cross-blending is proposed. A label cross-blending module is designed to adjust the proportion of labeled pixels while suppressing noisy and incorrect annotations in the generated training samples. In addition, a novel network architecture is developed to exploit high-level feature representations extracted by convolutional neural networks (CNNs), enabling more effective characterization of irregular targets and improved boundary delineation to enhance model robustness and segmentation consistency. Experimental results on the public MSCMR and ACDC datasets demonstrate the effectiveness of the proposed method, achieving Dice similarity coefficients of 86.8% and 87.1%, respectively.