摘要
目的:评估深度学习方法在鼻咽侧位X线片上自动测量儿童腺样体肥大的可行性、准确性和可靠性。方法:X线图像被人工标注并分为训练、验证和测试数据集。利用训练集和验证集对多分类U-Net和Res U-Net深度学习网络进行训练,然后分别用两种方法对测试集图像进行分割。将两种方法的分割结果进行比较,选择出最佳的分割方法,并利用Matlab所构建的测量模型自动测量腺样体/鼻咽腔(A/N)比值。将自动测量方法(AMS)测得的A/N值分别与由主任医师、主治医师、住院医师手工测量的A/N值进行比较。以主任医师的测量结果为标准,计算AMS、主治医师、住院医师的准确率。结果:与U-Net相比,Res U-Net在测试集有更好的分割性能,并获得了总体上与主治医师水平相当,但比住院医师更准确的A/N结果。在对正常、中度肥大和病理性肥大腺样体的分级中,AMS的准确率分别为93.75%、93.02%和96.00%,主治医师的准确率分别为100%、83.72%和96.00%,而住院医师的准确率分别为68.75%、69.77%和84.00%。统计学分析结果显示,在正常组、中度肥大组以及病理性肥大组中AMS与主任医师、主治医师的测量结果差异均无统计学意义(P值均>0.05);AMS与住院医师在正常组和中度肥大组中的测量结果差异有统计学意义(P<0.05),在病理性肥大组中差异无统计学意义(P>0.05)。AMS所需的测量时间分别为主任医师、主治医师和住院医师手工测量时间的22%、21%和18%,AMS与主任医师、主治医师、住院医师的测量时间比较差异均有统计学意义(P值均<0.05)。结论:基于Res U-Net的自动测量方法可以快速测量腺样体A/N比值,达到主治医师的测量水平,且能够提高住院医师的测量水平,辅助各级医师减少测量所需时间,有助于临床快速、精确诊断。
Objective:To evaluate the feasibility,accuracy and reliability of deep learning methods for automatic measurement of adenoid hypertrophy in children on nasopharyngeal lateral radiographs.Methods:X-ray images were manually annotated and divided into training,validation and test datasets.Training and validation groups were used to train the multi-class U-Net and Res U-Net deep learning image segmentation methods,and then both methods were used to segment the test set images.The test set results from both segmentation methods were compared to determine which one had the best performance,and the best one was used to automatically measure Adenoid/Nasopharyngeal(A/N)ratios via Matlab measuring model.The A/N ratios measured by Automation Measurement Solution(AMS)were compared with those manually measured by three different physicians:chief,attending and resident.The accuracy of measurement results of AMS,attending physicians and residents was calculated using the measurement results of chief physicians as the standard.Results:Res U-Net had better segmentation performance in the test set compared to U-Net,and obtained A/N results that were overall comparable to the level of attending physicians but more accurate than that of residents.In the classification of normal,moderate and pathological hypertrophic adenoids,the AMS had accuracy of 93.75%,93.02%and 96.00%,respectively,compared to 100%,83.72%and 96%for attending physicians,respectively,as well as 68.75%,69.77%and 84%for residents,respectively.Statistical analysis showed that there was no significant difference in the measurement results between AMS,chief physician and attending physician in the normal group,moderate hypertrophy group and pathological hypertrophy group(all P>0.05).There was a significant difference in the measurement results between AMS and residents in the normal and moderate hypertrophy groups(P<0.05),but no significant difference in the pathological hypertrophy group(P>0.05).The measurement time required for AMS was 22%,21%and 18%of the manual measurement time of chief physician,attending physician and resident,respectively.There were statistically significant differences in the measurement time between AMS and chief physician,attending physician and resident physician(all P<0.05).Conclusions:The AMS based on Res U-NET can quickly measure A/N ratio,which can reach the measurement level of attending physicians,and can improve the measurement level of residents,assist physicians at all levels to reduce the measurement time,and contribute to rapid and accurate clinical diagnosis.
作者
王军
何生
张智星
梁敏茜
姜增誉
李健丁
WANG Jun;HE Sheng;ZHANG Zhi-xing(Department of Radiology,The First Hospital of Shanxi Medical University,Taiyuan 030001,China)
出处
《放射学实践》
CSCD
北大核心
2022年第9期1143-1149,共7页
Radiologic Practice
