Knowledge Distillation-Based Model Lightweight for Medical Image Segmentation

Accurate delineation of anatomical structures via medical image segmentation constitutes a fundamental prerequisite for clinical diagnostics and therapeutic planning. However, the exorbitant parameter footprint in contemporary deep segmentation architectures imposes prohibitive demands on hardware resources, severely curtailing their viability on edge platforms with computational constraints. This research presents a knowledge distillation-driven framework for model compression, leveraging R2U-Net as the knowledge source and MobileNetV2 as the compact recipient. Through novel boundary-constrained distillation (BCD) incorporating both hard and soft boundary learning mechanisms, we facilitate efficient knowledge transfer while preserving segmentation fidelity. This methodology significantly curtails computational overhead and enhances operational efficiency in resource-limited clinical settings.