Objective:Using the deep learning method,iterative labeling training was carried out through the human-in-the-loop method to establish a pituitary segmentation model and realize the arti-ficial intelligence(AI)measurement of pituitary volume.Methods:1285 cases of brain 3D T1WI images were divided into groups of 5～15 years old,16～25 years old,26～50 years old,and 51～70 years old.80 subjects of each age group were randomly selected and divided into 4 batches for the test.Initially,3 subjects of each group were selected for radiologist pre-labeling of neurohypophysis and adenohy-pophysis,and input to the computer for learning to obtain the initial model.This model was then used to segment a batch of data to obtain the labeled neurohypophysis,adenohypophysis,and total pituitary volume.The segmentation results were manually calibrated by a radiologist,and the corresponding vol-ume data after calibration was obtained as the gold standard.The computer iteratively learned to opti-mize the model with the previous set of calibrated segmentation images,and then used the optimized model to segment and calibrate a new set of images.The above process was repeated,and finally the deep learning modeling was identified successful if the difference between the data before and after cal-ibration was not statistically significant(P＞0.05).Data were analyzed by paired t-test,Dice,and Spearman correlation analysis.Results:From the second batch,there was no statistically significant difference in neurohypophysis volume before and after calibration except for the 5～15 years old group,while the difference between adenohypophysis and total pituitary volume was statistically sig-nificant(P＜0.05).In the fourth batch,the neurohypophysis,adenohypophysis,and total pituitary vol-ume were more consistent before and after calibration across all age groups(P=0.137～0.928),the Dice value was greater than 0.941,and the Spearman correlation was greater than 0.969.Conclusion:Deep learning modeling through iterative training can build a human pituitary segmentation model and realize AI automatic measurement of pituitary volume.
关键词
垂体/深度学习/人在回路/自动分割/迭代式训练
Key words
Pituitary/Deep learning/Human in the loop/Automatic segmentation/Iterative training
维普
万方数据