Abstract：Objective Taking organ at risk (OAR) of lung cancer as an example, through the adjustment of fixed window width (WW)
and window level (WL) on CT images of thorax, to explore the influence of different WW/WL on the automatic delineation of OAR
of lung cancer based on deep learning. Methods The OAR (including left and right lungs, esophagus, spinal cord and heart) in the
2017 lung cancer OAR segmentation competition were automatically segmented by 2D-Unet. WW/WL adjustment was pretreated
before the training, that is, soft tissue window, lung window, mediastinal window, bone window and full window width were adjusted
for the CT images, and then the same conditions were used for training for each organ at risk. There were a total of 60 datasets in the
study, 48 of which were training sets, and the remaining 12 were testing sets. Dice similarity coefficient (DSC) and 95% hausdorff
distance (HD) were used as the evaluation criteria for the segmentation results of OAR. Statistical methods were determined by
Kruskal-Wallis H rank sum test or analysis of variance. Results Different WW/WL regulation had no significant effect on the DSC
values of left, right lung and spinal cord (P=0.057, 0.090, 0.894). The DSC values of esophagus and heart were significantly affected
(P<0.001). There was no significant difference in 95%HD value of spinal cord under different WW/WL (P=0.116). 95%HD values
of left and right lung, esophagus and heart under different WW/WL were statistically significant (P=0.005, 0.001, 0.007, <0.001).
Conclusion Different fixed WW/WL adjustments have different effects on different OAR, and the appropriate WW/WL should be
selected for automatic segmentation of CT images based on deep learning.
余行，何奕松，傅玉川. 不同固定窗宽/窗位调节在医学图像自动分割中的应用研究[J]. 中国医疗设备, 2022, 37(3): 75-78.
YU Hang, HE Yisong, FU Yuchuan. Research on the Application of Different Fixed Window Width and Window Level Adjustment
in Automatic Segmentation of Medical Images. China Medical Devices, 2022, 37(3): 75-78.
