摘要
目的 将改良Transformer模型应用于乳腺结节超声报告自主生成,并对其可行性进行初步探讨.方法 收集832例乳腺结节患者(共1284个结节)的超声图像构建BND数据集,引入一种改良Transformer模型对BND数据集进行智能分析,生成相应文本报告,并与Ensemble Model、SSD、R-FCN模型进行比较;同时引入LGK数据集,将改良Transformer模型与TieNet、Kerp、VTI、RNCM模型进行比较.采用BLEU评分评估各模型的性能.结果 在BND数据集中,改良模型的BLEU-1、BLEU-2、BLEU-3及BLEU-4评分分别为0.547、0.474、0.352、0.282,均高于Ensemble Model、SSD、R-FCN模型.在LGK数据集中,改良Transformer模型的BLEU-1、BLEU-2、BLEU-3及BLEU-4评分分别为0.579、0.391、0.288、0.152.结论 改良Transformer模型能够快速识别乳腺结节并自主生成标准报告,与Ensemble Model、SSD、R-FCN模型相比,获得了良好的BLEU评分,同时该模型在LGK数据集中BLEU评分也较高,表明改良Transformer模型具有较高的文本泛化性能.
Abstract
Objective To employ a modified Transformer model for intelligent generation of ultrasound reports on breast nodules,and to preliminarily explore the feasibility.Methods The ultrasound images of 832 patients with breast nodules(1284 nodules in total)were collected to construct the BND dataset.A modified Transformer model was introduced to intelligence analyze the BND dataset and generate the corresponding text report,which was compared with the Ensemble Model,SSD and R-FCN models,respectively.Moreover,the LGK dataset was introduced to compare the modified model with TieNet,Kerp,VTI,RNCM models,respectively.The performance of the models was evaluated by BLEU score.Results In the BND dataset,the BLEU-1,BLEU-2,BLEU-3 and BLEU-4 scores of the modified model were 0.547,0.474,0.352 and 0.282,respectively,which were higher than those in Ensemble Model,SSD and R-FCN models.In the LGK dataset,the BLEU-1,BLEU-2,BLEU-3 and BLEU-4 scores of the modified model were 0.579,0.391,0.288 and 0.152,respectively.Conclusion The modified Transformer model exhibits the ability to quickly identify breast nodules and generate standard reports independently.Compared with Ensemble Model,SSD and R-FCN models,it achieves a higher BLEU score.Furthermore,the modified model demonstrates exceptional performance on the LGK dataset,indicating it has strong text generalization capability.
NETL
NSTL
维普
万方数据