基于主题模型的胸部X光片诊断报告异常检测方法

An abnormal chest X-ray diagnostic report detection method based on topic model

胸部X光片是患者胸部检查的优先选择,对患者的诊断治疗起着重要的作用。医生依据自身的经验和习惯书写胸部X光片诊断报告,由于一些主观或者客观的原因,会开具一些影像描述与诊断结论不相符的异常诊断报告,因此对诊断报告进行异常检测有着重要的研究意义。胸片诊断报告未登录词多、数据高维稀疏,缺乏大量有效标注,传统方法检测异常胸片诊断报告效果不佳,为此,提出了一种基于主题模型的胸部X光片诊断报告异常检测方法。首先用双向LSTM-CRF模型结合诊断报告中的字符级特征,获取特定的医疗术语特征,解决诊断报告中未登录词多,描述自由的问题。然后依据领域知识和模板将诊断报告进行有效的特征扩展,缓解数据稀疏问题。最后用LDA模型判断诊断报告中影像描述与诊断结论特征是否匹配,检测出异常胸片诊断报告。实验结果表明,在阈值为2的情况下,异常检测的准确率为92.82%,召回率为69.54%,检测性能优于传统方法的。

Chest X ray is the preferred choice for patients’ chest examinations and plays an important role in the diagnosis and treatment of patients. Doctors write chest X-ray diagnostic reports based on their own experience and habits. For some subjective or objective reasons, they will issue some abnormal diagnostic reports that do not match the diagnostic conclusions. Therefore, it is of great significance to carry out abnormal detection of the diagnostic reports. Chest X-ray diagnostic reports have many unknown words and sparse high-dimensional data and lack of a lot of effective labeling. Traditional methods are ineffective in detecting abnormal chest X-ray diagnostic reports. Therefore, this paper proposes an abnormal chest X-ray diagnostic report detection method based on topic model. Firstly, the bidirectional LSTM-CRF model is used to combine the character-level features in the chest radiograph diagnosis reports to obtain the specific medical terminology features, so as to solve the problem that the diagnosis reports have many unknown words and are described freely. Secondly, based on domain knowledge and template, the chest X-ray diagnosis reports are extended effectively to alleviate the problem of data sparsity. Finally, the LDA model is used to determine whether the image description in the diagnosis reports match the characteristics of the diagnosis conclusion, so as to detect the abnormal chest X-ray diagnosis reports. Experiments show that the accuracy of abnormal detection is 92.82 and the recall rate is 69.54 when the threshold is 2. The proposal has higher abnormal detection performance than the traditional methods.