ScholarWorks@성균관대학교

초록

Accurate segmentation of pituitary adenoma (PA), pituitary gland (PG), and internal carotid artery (ICA) in contrast-enhanced T1-weighted (T1CE) coronal MRI is essential for diagnosis and treatment planning. A transformer-based segmentation model using Swin-Unet was proposed, and the effects of image preprocessing strategies and loss functions were investigated. A retrospective dataset of 255 patients was analyzed, with manual annotations serving as the ground truth. Preprocessing strategies including N4 bias correction, WhiteStrip and Z-score normalization, and resampling/cropping were evaluated. Several loss functions, including combo loss, Dice loss, and focal loss, were systematically compared. Swin-Unet was benchmarked against a classical U-Net baseline. Segmentation results revealed that Swin-Unet with resampling, cropping, and White-Strip normalization achieved the highest Dice similarity coefficients (DSCs) of 0.825 for PA, 0.565 for PG, and 0.716 for ICA. Swin-Unet demonstrated markedly improved segmentation performance, especially for PG, compared to U-Net (mean DSC = 0.613). These findings suggest that transformer-based architecture models, combined with appropriate preprocessing and loss optimization, can provide a promising framework for accurate and reliable segmentation of pituitary-related structures.

키워드