人工智能在乳腺超声报告生成中的应用

Application of artificial intelligence in breast ultrasound report generation

目的超声是中国女性乳腺检查的常用方法,提升超声医生临床诊断效率对乳腺筛查的全面普及意义重大。尽管医生目前有自己的模板生成超声报告,但从模板到最终形成报告,还有很多繁琐工作需要医生完成。本文尝试通过人工智能(artificial intelligence,AI)技术,辅助医生完成部分繁琐工作,提高医生完成乳腺超声报告的效率。方法首先基于乳腺影像报告和数据系统的词汇索引以及医生临床经验,归纳出结构化特征描述语句;然后基于卷积神经网络(convolutional neural networks,CNN)+Transformer模型获取患者的良恶性分类信息,基于Unet和冒泡排序模型自动标记、测量病灶最大径和计算纵横比;最后AI综合上述信息生成初步诊断报告,用以辅助医生生成最终报告。结果AI报告特征描述语句准确率达81.52%,病灶最大径相对误差为10.8%,纵横比准确率达到100%,提升医生撰写报告效率达到68.31%。结论该模型在保证准确率的情况下,能够有效减少医生撰写报告时间,为优化乳腺超声检查流程提供了技术基础。

Objective Ultrasound is a common method of breast screening for women in China.Improving the clinical diagnosis efficiency of ultrasound doctors is of great significance to the comprehensive popularization of breast screening.Although doctors currently have their own template to generate ultrasound reports,there are still a lot of cumbersome work to be completed by doctors from the template to the final report.This paper attempts to assist doctors to complete some cumbersome work through artificial intelligence(AI)technology,so as to improve the efficiency of doctors in completing breast ultrasound report.Methods Firstly,based on the vocabulary index of breast image report and data system and doctors′clinical experience,the structured feature description sentences were summarized.Then the patients were classified as benign and malignant based on CNN+Transformer model,and the maximum diameter and aspect ratio of lesions were automatically labeled and measured based on UNET and bubble sorting model.Finally,AI synthesized the above information to generate a preliminary diagnosis report to assist doctors in generating the final report.Results The accuracy rate of the AI report feature description sentence was 81.52%,the maximum diameter relative error was 10.8%,the accuracy rate of aspect ratio was 100%,and the efficiency of doctors in writing reports was improved by 68.31%.Conclusions This model can effectively reduce the time for doctors to write reports while ensuring accuracy,and provides a technical basis for optimizing breast ultrasound examination procedures.