Classifying abnormalities in computed tomography radiology reports with rule-based and natural language processing models

Purpose: When conducting machine learning algorithms on classification and detection of abnormalities for medical imaging, many researchers are faced with the problem that it is hard to get enough labeled data. This is especially difficult for modalities such as computed tomography (CT) with potentially 1000 or more slice images per case. To solve this problem, we plan to use machine learning algorithms to identify abnormalities within existing radiologist reports, thus creating case-level labels that may be used for weakly supervised training on the image data. We used a two-stage procedure to label the CT reports. In the first stage, a rule-based system labeled a smaller set of cases automatically with high accuracy. In the second stage, we developed machine learing algorithms using the labels from the rule-based system and word vectors learned without supervision from unlabeled CT reports. Method: In this study, we used approximately 24,000 CT reports from Duke University Health System. We initially focused on three organs, the lungs, liver/gallbladder, and kidneys. We first developed a rule-based system that can quickly identify certain types of abnormalities within CT reports with high accuracy. For each organ and disease combination, we produced several hundred cases with rule-based labels. These labels were combined with word vectors generated using word2vec from all the unlabeled reports to train two different machine learning algorithms: (a) average of word vectors merged by logistic regression, and (b) recurrent neural network (RNN). Result: Performance was evaluated by receiver operating characteristic (ROC) area under the curve (AUC) over an independent test set of 440 reports for which those organs were manually labeled as normal or abnormal by clinical experts. For lungs, the performance was 0.796 for average word vector and 0.827 for RNN. Liver performance was 0.683 for average word vector and 0.791 for RNN. For kidneys, it was 0.786 for average word vector and 0.928 for RNN. Conclusion: It is possible to label large numbers of cases automatically. These rule-based labels can then be used to build a classification model for large numbers of medical reports. With word2vec and other transfer learning techniques, we can get a good generalization performance.