Abstract
Objective: The purpose of our study was to evaluate the detection performance and interpretation of chest radiographs from Japanese subjects by physicians, including radiologists, thoracic physicians, and non-thoracic physicians, using deep-learning-based automatic detection (DLAD).
Materials and methods: In total, 49 radiographs with evidence of nodules, 54 with inflammation, and 91 normal chest radiographs were retrospectively collected. Radiograph classification and lesion-per-patient analysis were performed based on the probability values using the area under the curve. In addition, the lesion-per-patient accuracy was evaluated considering the sensitivity and specificity of the threshold. To study lung nodule detection in chest radiographs with and without DLAD, we retrospectively collected 60 normal chest radiographs and 51 chest radiographs showing nodules. Nine physicians evaluated the chest radiographs without and with DLAD, and the results were compared using jackknife free-response receiver operating characteristic (JAFROC) figures of merit.
Results: The per-patient analysis yielded an area under the receiver operating characteristics (AUROC) of 0.976. The per-annotation analysis produced a specificity of 82.35%, and the average false positive rate per patient was 0.13. According to the reader study, the JAFROC of all groups was 0.79 without DLAD and 0.82 with DLAD.
Conclusion: This study confirms that abnormalities can be detected with high accuracy in chest radiographs of Japanese patients, which is particularly significant because DLAD has been developed without the use of Japanese data. This suggests that DLAD may be useful for doctors globally.
