Automated segmentation of pituitary adenomas, pituitary gland, and internal carotid arteries on routine coronal contrast-enhanced T1-weighted MRI: a single-sequence feasibility study.

Introduction: Pituitary adenomas (PAs) alter the anatomy of the pituitary gland (PG) and internal carotid arteries (ICAs), requiring accurate segmentation for surgical planning and risk assessment. Multi-sequence magnetic resonance imaging (MRI)-based automatic segmentation methods have been employed but remain limited in routine clinical practice due to protocol inconsistencies. This study aimed to evaluate the feasibility and clinical relevance of Swin-Unet and nnU-Net for automated segmentation of the PA, PG, and ICA using only contrast-enhanced T1-weighted (T1CE) coronal MRI. Methods: A retrospective cohort of 255 patients with surgically confirmed PAs was analyzed. T1CE coronal MRI, acquired from heterogeneous scanners and protocols, was used for model development. Manual annotations of the PA, PG, and ICA by expert neurosurgeons served as ground truth. Two deep learning architectures were evaluated: 1) Swin-Unet, a transformer-based U-Net variant, and 2) nnU-Net, a self-configuring CNN framework. Models were trained using standard preprocessing and five-fold cross-validation, with ensemble predictions generated for final evaluation. Model performance was assessed using the Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), true-positive rate (TPR), and false-positive rate (FPR). Additionally, slice-level presence detection and per-patient processing time were evaluated. Results: Both models demonstrated comparable voxel-wise performance, with mean DSC = 0.70 (95% CI: 0.66-0.74) for Swin-Unet and 0.69 (95% CI: 0.65-0.73) for nnU-Net. Segmentation accuracy was lower for the PG compared to the PA and ICA. Boundary-based evaluation showed a mean HD95 of 5.21 mm (95% CI: 4.40-6.03) for Swin-Unet and 5.30 mm (95% CI: 4.44-6.16) for nnU-Net. Slice-level recognition yielded a mean F1-score of 0.89 ± 0.09 for Swin-Unet and 0.87 ± 0.08 for nnU-Net across the PA, PG, and ICA, respectively. The mean computation time was 78.4 seconds per patient (95% CI: 61.5-95.5), and the segmentation outputs facilitated approximate 3D reconstructions for qualitative assessment. Discussion: Our findings demonstrate the feasibility of automated segmentation of the PA, PG, and ICA using single-sequence T1CE coronal MRI. While voxel-level accuracy was modest, the combination of reliable slice-level recognition, rapid processing, and 3D visualization suggests potential utility as an adjunct tool for radiologic assessment and surgical planning.