Comparison of Chest Radiograph Captions Based on Natural Language Processing vs Completed by Radiologists

Abstract

Artificial intelligence (AI) can interpret abnormal signs in chest radiography (CXR) and generate captions, but a prospective study is needed to examine its practical value. To prospectively compare natural language processing (NLP)-generated CXR captions and the diagnostic findings of radiologists. A multicenter diagnostic study was conducted. The training data set included CXR images and reports retrospectively collected from February 1, 2014, to February 28, 2018. The retrospective test data set included consecutive images and reports from April 1 to July 31, 2019. The prospective test data set included consecutive images and reports from May 1 to September 30, 2021. A bidirectional encoder representation from a transformers model was used to extract language entities and relationships from unstructured CXR reports to establish 23 labels of abnormal signs to train convolutional neural networks. The participants in the prospective test group were randomly assigned to 1 of 3 different caption generation models: a normal template, NLP-generated captions, and rule-based captions based on convolutional neural networks. For each case, a resident drafted the report based on the randomly assigned captions and an experienced radiologist finalized the report blinded to the original captions. A total of 21 residents and 19 radiologists were involved. Time to write reports based on different caption generation models. The training data set consisted of 74 082 cases (39 254 [53.0%] women; mean [SD] age, 50.0 [17.1] years). In the retrospective (n = 8126; 4345 [53.5%] women; mean [SD] age, 47.9 [15.9] years) and prospective (n = 5091; 2416 [47.5%] women; mean [SD] age, 45.1 [15.6] years) test data sets, the mean (SD) area under the curve of abnormal signs was 0.87 (0.11) in the retrospective data set and 0.84 (0.09) in the prospective data set. The residents' mean (SD) reporting time using the NLP-generated model was 283 (37) seconds-significantly shorter than the normal template (347 [58] seconds; P < .001) and the rule-based model (296 [46] seconds; P < .001). The NLP-generated captions showed the highest similarity to the final reports with a mean (SD) bilingual evaluation understudy score of 0.69 (0.24)-significantly higher than the normal template (0.37 [0.09]; P < .001) and the rule-based model (0.57 [0.19]; P < .001). In this diagnostic study of NLP-generated CXR captions, prior information provided by NLP was associated with greater efficiency in the reporting process, while maintaining good consistency with the findings of radiologists.