Abstract
Despite the use of T1CE coronal MRI in PA diagnosis, the effectiveness of automatic segmentation remains suboptimal, with research predominantly focused on convolutional neural networks. We aim To develop a transformer-based model for the automatic segmentation of pituitary adenoma (PA), pituitary gland (PG), and internal carotid artery (ICA) in contrast-enhanced T1-weighted (T1CE) coronal MR images. Patients with PA were retrospectively selected for inclusion in the training and test datasets using electronic medical and radiological records. Pretreatment T1CE coronal MR images were collected, annotated, and used as ground truth. A Swin-Unet-based transformer model was developed for segmentation, incorporating preprocessing techniques and loss functions to address class imbalance. To evaluate performance, the nnU-Net V2 model was trained using the same training dataset and compared against the Swin-Unet model. A total of 255 patients with PA were analyzed for model training and evaluation. The Swin-Unet model demonstrated superior performance compared to raw images, with Dice similarity coefficients (DSCs) of 0.825, 0.565, and 0.716 for PA, PG, and ICA, respectively (p=1.000, <0.01, and 0.425). Combo loss revealed the best segmentation performance, with the highest DSC of 0.702, a 95th percentile Hausdorff distance (95HD) of 4.846 mm, a true-positive rate (TPR) of 0.706, and a false-positive rate (FPR) of 0.002 for the mean of three classes. The nnU-Net V2 model exhibited comparable performance to Swin-Unet, with DSC of 0.691, 95HD of 3.686 mm, TPR of 0.775 (p < 0.001), and FPR of 0.002 for the mean of three classes. A novel transformer-based segmentation model using Swin-Unet was introduced for the automated detection of PA, PG, and ICA in T1CE coronal MRI.