Akademik Veri Yönetim Sistemi
Annals of Computer Science and Intelligence Systems, cilt.45, ss.72-76, 2025 (Hakemli Dergi)
Brain extraction, or skull stripping, is a crucial preprocessing step in magnetic resonance imaging (MRI), isolating brain tissue from surrounding structures like the skull and scalp. However, existing methods have limitations, such as parameter sensitivity in traditional approaches and computational complexity in advanced deep learning architectures. This study proposes a knowledge distillation framework utilizing two UNet++ models—a high-capacity teacher network and an efficient student network—for 3D brain extraction tasks. The teacher network generates detailed grayscale brain predictions, capturing subtle intensity transitions and anatomical boundaries. The student network learns to produce precise binary segmentation masks from the teacher’s feature representations, guided by a hybrid loss function combining Dice, Structural Similarity Index Measure (SSIM), and Mean Squared Error (MSE). Evaluations conducted on T1-weighted, T2-weighted, and proton-density weighted MRI images from the IXI dataset demonstrated the student model’s superior performance, achieving a Dice coefficient of 0.97857. These findings suggest that the proposed framework may offer a practical and accurate solution for brain extraction in diverse medical imaging scenarios.