Pancreatic cancer stands out as one of the most challenging and lethal cancer to treat, primarily due to its late - stage detection and the limited therapeutic options. Pancreatic cancer often goes undetected until late stages because of the pancreas’s irregular shape, small size, and varying texture across patients. For early diagnosis and efficient treatment planning, pancreatic cancers in radiological images must be precisely segmented. To address the anatomical variability of the pancreas across slices (intra class heterogeneity) and the minimal contrast distinguishing the pancreas from anatomical structures (inter-class ambiguity), proposed a hybrid 2D-3D U-Net framework. This study introduces a hybrid deep learning technique for automated 2D-3D volumetric segmentation of pancreas in computed tomography (CT) images. This combination strategy enables the network to captures the both global as well as local textures from CT images. The model was trained using manually annotated ground truth masks from the NIH Pancreas-CT dataset, comprising 18,942 DICOM images from 80 subjects. The proposed multi-stage pipeline incorporates a one-cycle learning rate policy, and employ a Tversky loss function to mitigate class imbalance. To evaluate the segmentation performances, standard metrics such as Dice Similarity Coefficient (DSC) is mainly focused. Experimental outcomes show that the suggested method accomplishes accurate and reliable pancreas segmentation, highlighting its promise for integration into computer-aided diagnostic tools improve pancreas segmentation supporting its integration into clinical decision-support systems.