调整窗宽/窗位对卷积神经网络模型自动筛选胰腺肿瘤CT图像性能的影响

目的 观察调整窗宽/宽位对卷积神经网络（CNN）模型自动筛选胰腺肿瘤CT图像性能的影响。方法 按6 ∶ 4比例将医学图像分割十项全能挑战赛中的胰腺CT数据集（包含281例胰腺肿瘤CT图像共26 719幅）数据分为训练集和测试集；训练集共15 346幅，33.96%（5 212/15 346）存在胰腺肿瘤，测试集共11 373幅，34.26%（3 896/11 373）存在胰腺肿瘤。采用调窗方法，分别设置窗位为30、40及50 HU，以120、300和500 HU为窗宽范围端点，经组合配对形成9组参数，调整原始数据集窗宽/窗位，将原始数据直接映射到0~255像素灰度值，共得到10组数据，分别将其导入Alexnet-V1、Alexnet-V2、Resnet-V1及Resnet-V2共4个CNN模型进行训练，于相应测试集中筛选胰腺肿瘤图像，以准确率（Acc）、敏感度（Sen）及特异度（Spe）评价其效能。结果 不同CNN模型用于筛选测试集无调窗数据胰腺肿瘤图像的Acc均为34.26%，Sen均为100%，Spe均为0。调整测试集图像窗位均为30 HU时，CNN模型对于窗宽300 HU图像的Acc及Spe最高，分别为（84.17±1.89）%及（77.91±1.96）%；窗位为40 HU及50 HU时，CNN模型在窗宽300 HU图像中的Acc、Sen及Spe分别为（85.98±2.66）%、（97.19±1.41）%及（82.12±3.44）%和（84.29±2.38）%、（97.68±1.65）%及（77.52±5.35）%，均高于120 HU和500 HU。结论 通过调整窗宽/窗位可在CNN自动筛选胰腺肿瘤图像预处理过程中初步排除冗余信息；合理设置窗宽/窗位能有效提高CNN模型的筛选性能。

Impact of adjusting window width and window level on performance of convolutional neural network model for automatic screening CT images including pancreatic neoplasms

Objective To observe the impact of adjusting window width and window level on performance of convolutional neural network (CNN) model for automatic screening CT images including pancreatic neoplasms. Methods Data of pancreas CT data set (including 26 719 CT images of 281 cases of pancreatic neoplasms) in Medical Segmentation Decathathon were divided into training set and test set in the ratio of 6:4. There were 15 346 images in training set, 33.96% (5 212/15 346) had pancreatic neoplasms, and 11 373 images in test set, 34.26% (3 896/11 373) had pancreatic neoplasms. Using clinical window adjustment method, set window level as 30, 40 and 50 HU, and selected 120, 300 and 500 HU as the end points of window width range, respectively, 9 groups of parameters were obtained for adjusting window width/window level of the original data set. Moreover, the original data were directly mapped to the gray value of 0-255 pixels for processing, then totally 10 data sets were obtained. The images of 10 groups of training set were imported into 4 CNN models of Alexnet-V1, Alexnet-V2, Resnet-V1 and Resnet-V2 models for training, then applied to screen images including pancreatic neoplasms in the corresponding test set, and screening efficacy was evaluated with accuracy (Acc), sensitivity (Sen) and specificity (Spe). Results Acc, Sen and Spe of different CNN models were all 34.26%, 100% and 0 for screening images including pancreatic neoplasms in untuned window data of test set. When window level of test set was 30 HU, Acc and Spe of CNN model on images of window width 300 HU were the highest, which was (84.17±1.89)% and (77.91±1.96)%, respectively. When window level of test set was adjusted to 40 HU and 50 HU, Acc, Sen and Spe of CNN model on images with window width of 300 HU was (85.98±2.66)%, (97.19±1.41)% and (82.12±3.44)%, respectively, and (84.29±2.38)%, (97.68±1.65)% and (77.52±5.35)%, respectively, all higher than those when window width was 120 HU and 500 HU. Conclusion Adjusting window width/window level could preliminarily eliminate redundant information in preprocessing of CNN model automatic screening CT images including pancreatic neoplasms. Reasonable window width/window level parameters could effectively improve screening performance of CNN model.