Research On MRI Image Segmentation Of Pituitary Adenoma Based On Deep Learning

Objective:Image segmentation is one of the key steps in the field of medical image analysis.The segmentation efficiency and precision of results are directly related to the timeliness and accuracy of clinical treatment and scientific research.With the progress of deep neural network architecture and the improvement of medical image segmentation requirements,Deep learning methods are widely used in various medical image processing tasks,including tumor segmentation.At present,there are few reports on the segmentation of pituitary adenoma MRI images using deep learning methods at home and abroad.This study intends to explore the effect of deep learning model based on convolutional neural network on semantic segmentation of pituitary adenoma and sellar tissue（pituitary,optic nerve,internal carotid artery and cavernous sinus）on 3D-CUBE T1 WI contrast-enhanced MRI images,The segmentation results of the deep learning model are applied to the 3D reconstruction of the sellar region to improve the segmentation efficiency of MRI images of pituitary adenomas and promote its transformation to clinical application.Data and methods:A total of 72 patients（5816 coronal and sagittal high-resolution MRI images）who met the research conditions underwent pituitary MRI 3D-CUBE T1 WI contrast-enhanced examination in our hospital from January 2019 to December 2021 were collected.The average age of the patients was 24 ± 10.5 years,and the average value of the maximum diameter of tumor tissue was 16 ± 4.2mm;There were 58 cases of prolactinoma,7 cases of growth hormone tumor,5 cases of adrenocorticotropic hormone tumor and 2 other cases.After image clipping,preprocessing and data enhancement,a 3D dataset containing MRI volume images of 1008 pituitary adenomas was constructed,80% of which was used as the training dataset and 20% as the test dataset.It was input into the deep learning model based on four different 3D CNN architectures（U-Net3 D,Dense-Net 3D,V-Net and Dense-Voxel-Net）for training and testing,The segmentation effects of each network on pituitary adenoma and adjacent sellar tissue were compared,and the segmentation results were applied to 3D reconstruction of sellar region.Result:Accuracy results: the depth neural network showed good learning ability in the MRI image segmentation task of pituitary adenoma.The average dice coefficients of the depth learning models based on four different 3D CNNs on the test dataset were84.83%（U-Net 3D）,85.41%（Dense-Net 3D）,81.44%（V-Net）and 83.42%（Dense-Voxel-Net）respectively;The model performs well in the segmentation of tumor tissue,optic nerve and brain tissue and bilateral internal carotid arteries,reaching 91.53%（U-Net 3D）,90.41%（Dense-Net 3D）and 90.80%（U-Net 3D）respectively,while the segmentation effect of normal pituitary tissue and bilateral cavernous sinus is poor,reaching54.22%（U-Net 3D）and 39.78%（Dense-Voxel-Net）respectively.Training time and video memory occupation: the training time of the four networks and the use of GPU video memory are 9h 26 min / 16.2 GB（U-Net 3D）,12 h 15min / 19.0GB（Dense-Net 3D）,11 h 33min / 20.5 GB（V-Net）and 12 h 30min / 21.0 GB（Dense-Voxel-Net）respectively.Weight parameter volume and unit image time consumption: the weight parameter volume of each network and the time required to process a single image are 6.8 MB/ 0.52s（U-Net 3D）,6.02 MB / 0.72s（Dense-Net 3D）,174 MB / 1.63s（V-Net）and 6.88 MB / 0.75s（Dense-Voxel-Net）respectively.conclusionThe 3D CNN has a good effect on the segmentation of pituitary adenoma and sellar tissue on 3D-CUBE T1 WI contrast-enhanced MRI images,mainly in the segmentation of tumor tissue,optic nerve and brain tissue and bilateral internal carotid arteries.The verification results of the test dataset show that the model has certain robustness and generalization ability,It proves the effectiveness of deep learning method in pituitary adenoma segmentation task.Visualization and 3D reconstruction of segmentation results can meet some clinical and scientific research needs,and provide valuable information for clinical treatment,preoperative evaluation and prognosis evaluation of patients with pituitary adenoma under the background of artificial intelligence and precision medicine.