Transformer-stage based segmentation of empty sella and peripheral arteries

Empty sella is an incipient disease inducing various critical illnesses, such as visual disturbance, intrasellar tumors, and idiopathic intracranial hypertension. The pituitary stalk and the internal carotid artery are important anatomic markers in imaging localization. Diagnosing the empty sella with automatic approach can be more accurate and efficient, which is also the premise for area calculation and researching the relationship between this artery and peripheral arteries. Previous studies primarily focus on the case reports and late-stage symptoms of this disease, without mentioning its influence. Besides, since empty sella overlaps with pituitary stalk and has different relative positions to the internal carotid artery, it is difficult to separate them. To solve these problems, we proposed a transformer-stage based multi-modality method involving spatial, boundary, and semantic attention modules, and constructed a medical image dataset for empty sella, including 288 samples in 4 modalities, which was evaluated by various experiments. Results show that the proposed network, with the mIoU of 0.8944, is superior to SOTA networks, which proves the excellent recognition potential of the transformer and multiple modalities in this dataset.