Journal of Advances in Mathematics and Computer Science

Pancreas segmentation from computed tomography (CT) images is challenging because of anatomical variability, irregular morphology, and the complexity of surrounding abdominal structures. Although deep neural networks can provide strong segmentation performance, their computational demands and prolonged training can limit use in resource-constrained settings. This study evaluates Structured Knowledge Distillation via Collaborative Learning (SKDCL), a teacher-free, single-stage framework that integrates collaborative learning with structured distillation for lightweight pancreas segmentation. Six architectures—UNet, PSPNet, DeepLabv3+, ERFNet, ESPNet, and ENet—were evaluated in paired combinations under consistent experimental settings. The NIH Pancreas-CT dataset comprised 6,882 abdominal CT images from 82 patients and was divided into training, validation, and test sets in proportions of 70%, 15%, and 15%, respectively. Structured distillation incorporated region-level and prediction-level information to strengthen spatial representation learning, while performance was assessed using the Dice similarity coefficient (DSC) and total training time. Collaborative learning improved segmentation performance relative to independent model training, and structured distillation produced further gains across the evaluated combinations. The highest DSC was 93.56% for ESPNet, while ENet achieved 92.83% in a structured collaborative pairing. These findings show that SKDCL can support competitive segmentation performance in lightweight two-dimensional models without requiring a pre-trained teacher network. The framework therefore offers a computationally efficient approach to pancreas CT segmentation, although validation on additional datasets and extension to three-dimensional architectures remain necessary.