Automatic pituitary adenoma segmentation and identification of cavernous sinus invasion via multitask learning

This study aimed to develop a multitask deep learning model for pituitary macroadenoma (PMA) segmentation and identification of cavernous sinus (CS) invasion.

A total of 926 patients with PMAs (n=816 from Institution 1 for model training and n=110 from Institution 2 for model validation) were included. Manual segmentation for tumor and bilateral internal carotid arteries was conducted on coronal contrast-enhanced T1-weighted imaging images using ITK-SNAP. Two performing neurosurgeons evaluated the CS invasion during the operation. A Multi-Task Multiaxis-Attention UNet (MTMAU-Net) framework that combined segmentation and CS invasion classification tasks was proposed. Dice similarity coefficient and 95% Hausdorff distance (HD95) were used to evaluate the segmentation performance. Accuracy and area under the curve (AUC) were used to assess the classification performance.

Compared with several single-task models, MTMAU-Net model achieved higher Dice coefficient (90.84±0.55%) and lower HD95 values (2.13±0.37 mm) in segmentation tasks and performed better in the classification task (accuracy: 88.05±2.84%, AUC: 0.89±0.08). Compared with the Knosp grading system, this model showed overall consistent performance (accuracy: 88.05% vs 87.68%) and outperformed human grading evaluation at Knosp grade 3 (83.33% vs 57.62%). MTMAU-Net showed great capability of generalization in both segmentation (Dice coefficient: 83.71±5.93%) and classification tasks (accuracy: 84.55%, AUC: 0.87) in the validation cohort.

MTMAU-Net outperformed single-task models in PMA segmentation and CS invasion classification and showed advantages over the Knosp grading system in identifying CS invasion, which might guide clinicians to make appropriate surgical strategies.

This study is registered at the Chinese Clinical Trial Registration Center with registration number ChiCTR210047614.