<p>Pancreatic segmentation in abdominal CT imaging remains challenging because the pancreas is small, weakly contrasted against surrounding tissues, and difficult to annotate at scale. To address these issues, we propose MPC-CAU-Net, a pancreas-oriented framework that combines Inter-layer Positional Contrastive Learning (IPCL), CBAM-enhanced skip connections, and multi-view fusion. IPCL exploits normalized slice-wise relative depth to form contrastive pairs within each view-specific volume, enabling the encoder to learn inter-slice contextual continuity as a weak positional prior. CBAM is introduced into the skip pathways to suppress background interference and strengthen faint boundary textures. The axial, coronal, and sagittal 2D predictions are finally fused by voxel-wise majority voting. Experiments on the NIH pancreas CT dataset show that MPC-CAU-Net achieves a Dice score of 82.94% and an IoU of 71.11%, outperforming the single-view PC-CAU-Net by about 2.09 and 2.94 percentage points, respectively. Under self-supervised pretraining, IPCL improves Dice by about 2.01 points over SimCLR and 1.59 points over GCL at the 100% annotation setting. These results support the effectiveness of combining positional pretraining, attention-guided refinement, and cross-view complementarity for pancreas CT segmentation. Since the current experiments are limited to a single public CT dataset, cross-center and cross-modality applicability should be regarded as a promising but not yet directly validated direction.</p>