基于深度学习的放大蓝激光成像和放大窄带成像模式下早期胃癌识别模型的诊断效果对比

Comparison of the diagnostic effect of early gastric cancer between magnifying blue laser imaging model and magnifying narrow-band imaging model based on deep learning

目的研制基于深度卷积神经网络的放大蓝激光成像(ME-BLI)和放大窄带成像(ME-NBI)模式下早期胃癌识别系统,比较2种模式下模型的性能差异,并探讨模型训练方式对准确度的影响。方法回顾性收集ME-BLI和ME-NBI下的胃良性病灶和早期胃癌的内镜图像,共收集5个数据集和3个测试集:数据集1包括2024张非癌和452张早期胃癌ME-BLI图片,数据集2包括2024张非癌和452张早期胃癌ME-NBI图片,数据集3是数据集1和2的合集(共4048张非癌、904张早期胃癌ME-BLI和ME-NBI图片),数据集4在数据集2的基础上增加62张非癌和2305张早期胃癌ME-NBI图片(共2086张非癌和2757张早期胃癌ME-NBI图片),数据集5在数据集3的基础上增加62张非癌和2305张早期胃癌ME-NBI图片(共4110张非癌、3209张早期胃癌ME-BLI和ME-NBI图片);测试集A包括422张非癌和197张早期胃癌ME-BLI图片,测试集B包括422张非癌和197张早期胃癌ME-NBI图片,测试集C是测试集A和测试集B的合集(共844张非癌、394张早期胃癌ME-BLI和ME-NBI图片)。根据5个数据集分别构建5个模型,分别评估其在3个测试集中的表现。通过以病灶为单位的视频测试,比较ME-NBI和ME-BLI模式下早期胃癌识别模型在临床环境下的性能差异,并与4名资深内镜医师进行比较。主要终点指标为早期胃癌的诊断准确度、灵敏度和特异度。采用卡方检验进行统计学分析。结果模型1在测试集A的表现最佳,准确度、灵敏度、特异度分别为76.90%(476/619)、63.96%(126/197)、82.94%(350/422);模型2在测试集B的表现最佳,准确度、灵敏度和特异度分别为86.75%(537/619)、92.89%(183/197)、83.89%(354/422);模型3在测试集B中的表现最佳,准确度、灵敏度和特异度分别为86.91%(538/619)、84.26%(166/197)、88.15%(372/422);模型4在测试集B的表现最佳,准确度、灵敏度和特异度分别为85.46%(529/619)、95.43%(188/197)、80.81%(341/422);模型5在测试集B的表现最佳,准确度、灵敏度和特异度分别为83.52%(517/619)、96.95%(191/197)、77.25%(326/422)。根据图片识别早期胃癌,模型2~5的准确度均高于模型1,差异均有统计学意义(χ^(2)=147.90、149.67、134.20、115.30,P均<0.01);模型2和3的灵敏度和特异度均高于模型1,模型2的特异度低于模型3,差异均有统计学意义(χ^(2)=131.65、64.15、207.60、262.03、96.73,P均<0.01);模型4和5的灵敏度均高于模型1~3,模型4和5的特异度均低于模型1~3,差异均有统计学意义(χ^(2)=151.16、165.49、71.35、112.47、132.62、153.14,176.93、74.62、14.09、15.47、6.02、5.80,P均<0.05)。以病灶为单位的视频测试结果显示,医师1~4的平均准确度为68.16%;模型1~5的准确度分别为69.47%(66/95)、69.47%(66/95)、70.53%(67/95)、76.84%(73/95)和80.00%(76/95)。模型1~5之间、模型1~5与医师1~4之间的准确度比较差异均无统计学意义(P均>0.05)。结论基于深度学习的ME-BLI早期胃癌识别模型具有较好的准确度,但诊断效果略差于ME-NBI模型;ME-NBI联合ME-BLI早期胃癌识别模型的诊断效果优于单独模式下的模型;增加ME-NBI图片数量,尤其是早期胃癌图片,可以提高ME-NBI模型的灵敏度,但特异度有所下降。

Objective To develop early gastric cancer(EGC)detection system of magnifying blue laser imaging(ME-BLI)model and magnifying narrow-band imaging(ME-NBI)model based on deep convolutional neural network,to compare the performance differences of the two models and to explore the effects of training methods on the accuracy.Methods The images of benign gastric lesions and EGC under ME-BLI and ME-NBI were respectively collected.A total of five data sets and three test sets were collected.Data set 1 included 2024 noncancerous lesions and 452 EGC images under ME-BLI.Data set 2 included 2024 noncancerous lesions and 452 EGC images under ME-NBI.Data set 3 was the combination of data set 1 and 2(a total of 4048 noncancerous lesions and 904 EGC images under ME-BLI and ME-NBI).Data set 4:on the basis of data set 2,another 62 noncancerous lesions and 2305 EGC images under ME-NBI were added(2086 noncancerous lesions and 2757 EGC images under ME-NBI).Data set 5:on the basis of data set 3,another 62 noncancerous lesions and 2305 EGC images under ME-NBI were added(4110 noncancerous lesions and 3209 EGC images under ME-NBI and ME-BLI).Test set A included 422 noncancerous lesions and 197 EGC images under ME-BLI.Test set B included 422 noncancerous lesions and 197 EGC images under ME-NBI.Test set C was the combination of test set A and B(844 noncancerous and 394 EGC images under ME-BLI and ME-NBI).Five models were constructed according to these five data sets respectively and their performance was evaluated in the three test sets.Per-lesion video was collected and used to compare the performance of deep convolutional neural network models under ME-BLI and ME-NBI for the detection of EGC in clinical environment,and compared with four senior endoscopy doctors.The primary endpoint was the diagnostic accuracy of EGG,sensitivity and specificity.Chi-square test was used for statistical analysis.Results The performance of model 1 was the best in test set A with the accuracy,sensitivity and specificity of 76.90%(476/619),63.96%(126/197)and 82.94%(350/422),respectively.The performance of model 2 was the best in test set B with the accuracy,sensitivity and specificity of 86.75%(537/619),92.89%(183/197)and 83.89%(354/422),respectively.The performance of model 3 was the best in test set B with the accuracy,sensitivity and specificity of 86.91%(538/619),84.26%(166/197)and 88.15%(372/422),respectively.The performance of model 4 was the best in test set B with the accuracy,sensitivity and specificity of 85.46%(529/619),95.43%(188/197)and 80.81%(341/422),respectively.The performance of model 5 was the best in test set B,with the accuracy,sensitivity and specificity of 83.52%(517/619),96.95%(191/197)and 77.25%(326/422),respectively.In terms of image recognition of EGC,the accuracy of models 2 to 5 was higher than that of model 1,and the differences were statistically significant(χ^(2)=147.90,149.67,134.20 and 115.30,all P<0.01).The sensitivity and specificity of models 2 and 3 were higher than those of model 1,the specificity of model 2 was lower than that of model 3,and the differences were statistically significant(χ^(2)=131.65,64.15,207.60,262.03 and 96.73,all P<0.01).The sensitivity of models 4 and 5 was higher than those of models 1 to 3,and the specificity of models 4 and 5 was lower than those of models 1 to 3,and the differences were statistically significant(χ^(2)=151.16,165.49,71.35,112.47,132.62,153.14,176.93,74.62,14.09,15.47,6.02 and 5.80,all P<0.05).The results of video test based on lesion showed that the average accuracy of doctors 1 to 4 was 68.16%.And the accuracy of models 1 to 5 was 69.47%(66/95),69.47%(66/95),70.53%(67/95),76.84%(73/95)and 80.00%(76/95),respectively.There were no significant differences in the accuracy among models 1 to 5 and between models 1 to 5 and doctors 1 to 4(all P>0.05).Conclusions ME-BLI EGC recognition model based on deep learning has good accuracy,but the diagnostic effecacy is sligntly worse than that of ME-NBI model.The effects of EGC recognition model of ME-NBI combined with ME-BLI is better than that of a single model.A more sensitive ME-NBI model can be obtained by increasing the number of ME-NBI images,especially the images of EGG,but the specificity is worse.