| BackgroundPituitary adenomas(PAs)are benign tumors of the pituitary gland.PAs are classified by their size and whether or not they secrete hormones.Magnetic resonance imaging(MRI)scans are the most commonly used for clinical diagnosis and treatment of PA.Endoscopic transnasal transsphenoidal surgery(ETTS)is the most common method for removing PA since it is minimally invasive and has a short hospital stay.Among the developments in minimally invasive and individualized PA surgery,the use of artificial intelligence for preoperative assessment and surgical planning has received increasing attention.ObjectiveTraining deep learning models using MRI images for automatic PA segmentation,preoperative prediction of Ki67 status,and intraoperative cerebrospinal fluid(CSF)leak.In addition,analysis of giant pituitary adenoma(GPA)morphology using a novel 3D cube system and evaluation surgical risk of ETTS.Methods:(1)MRI images from 243 PA patients were collected,and nnU-Net architecture was used to train deep learning models for PA semantic segmentation.Dice similarity coefficient(DSC)was used to test model performance.Model 1 was trained with TICE images from all kinds of PA,and Model 2 was trained with primary nonfunctional pituitary adenomas(pNFPA)images only.Model performance was compared to investigate cost-effective ways to train deep learning models for clinical application.(2)MRI images and PA pathology results of 243 PA patients were reviewed.MRI images were labeled with Ki67 index<3%or Ki67 index>3%,and a U-Net architecture was used to train T1CE,T2W and duo-modal(T1CE+T2W)models.Model performance was evaluated with prediction accuracy.Performance of the three models were evaluated and compared on both validation and test datasets.(3)A total of 298 PA patients were included in this study.Patient MRI images and surgical records were reviewed and intraoperative CSF leaks was confirmed with intraoperative observations.A U-Net architecture was used to train three models with T1CE,T2W,and duo-modal images respectively.Model performance was evaluated with prediction accuracy.Performance of the three models were tested and compared on validation dataset.(4)A total of 39 GPA patients were included in the study.Automatic segmentation of GPA on T1CE was achieved by the nnU-Net model.Max diameter and volume of GPAs were measured.Three-dimensional(3D)cube system based on three floors,three columns,and three rows has been used to evaluate tumor morphology and difficulty in resecting tumors.Results:(1)On the validation dataset,both models performed well(DSC>0.8)for PAs with volume>1000mm3,but unsatisfactorily(DSC<0.5)for PAs<1000mm3.On the testing dataset,the mean DSC values of both models achieved 0.72.Model 2 trained with fewer samples was more cost-effective in clinical practice.(2)Three deep learning models were trained with T1CE,T2W,and duo-modal images for the prediction of PA Ki67 status.On the validation dataset,both T2W model and duo-modal model achieved the same accuracy of 86.75%.On the testing dataset,the duo-modal model outperformed the other two models,the prediction accuracy of the duo-modal model achieved 88.89%when single paired images as input data.(3)Three deep learning models were trained with T1CE,T2W,and duo-modal images for the prediction of intraoperative CSF leaks.Duo-modal model outperformed the T1CE and T2W models.The prediction accuracy,sensitivity accuracy,and specificity accuracy of the duo-modal model were 89%,88.5%,and 90.32%respectively.(4)A total of 39 GPA patients were included in this study,seven of whom underwent transcranial approach and 32 of whom underwent transsphenoidal approach.In different surgical approaches,different GPA morphology patterns were observed.ETTS is appropriate for tumors located in lower subcubes,whereas transcranial is appropriate for tumors extending into the higher subcubes.Conclusions:Automatic PA segmentation on T1CE was achieved with nnU-Net deep learning model.The mean DSC value achieved over 0.85 for the 75%of the PAs in clinical application.Preoperative prediction of Ki67 status and intraoperative CSF leaks were realized on deep learning models using MRI images,which provided more information for the treatment decision-making and prognostic evaluation.3D cube system based on PA 3D imaging and anatomic landmarks provide a new qualitative and quantitative method for evaluation of surgical risk. |
