Diagnostic performance of artificial intelligence model for pneumonia from chest radiography

OBJECTIVE: The chest X-ray (CXR) is the most readily available and common imaging modality for the assessment of pneumonia. However, detecting pneumonia from chest radiography is a challenging task, even for experienced radiologists. An artificial intelligence (AI) model might help to diagnose pneumonia from CXR more quickly and accurately. We aim to develop an AI model for pneumonia from CXR images and to evaluate diagnostic performance with external dataset. METHODS: To train the pneumonia model, a total of 157,016 CXR images from the National Institutes of Health (NIH) and the Korean National Tuberculosis Association (KNTA) were used (normal vs. pneumonia = 120,722 vs.36,294). An ensemble model of two neural networks with DenseNet classifies each CXR image into pneumonia or not. To test the accuracy of the models, a separate external dataset of pneumonia CXR images (n = 212) from a tertiary university hospital (Gachon University Gil Medical Center GUGMC, Incheon, South Korea) was used; the diagnosis of pneumonia was based on both the chest CT findings and clinical information, and the performance evaluated using the area under the receiver operating characteristic curve (AUC). Moreover, we tested the change of the AI probability score for pneumonia using the follow-up CXR images (7 days after the diagnosis of pneumonia, n = 100). RESULTS: When the probability scores of the models that have a threshold of 0.5 for pneumonia, two models (models 1 and 4) having different pre-processing parameters on the histogram equalization distribution showed best AUC performances of 0.973 and 0.960, respectively. As expected, the ensemble model of these two models performed better than each of the classification models with 0.983 AUC. Furthermore, the AI probability score change for pneumonia showed a significant difference between improved cases and aggravated cases (Δ = -0.06 ± 0.14 vs. 0.06 ± 0.09, for 85 improved cases and 15 aggravated cases, respectively, P = 0.001) for CXR taken as a 7-day follow-up. CONCLUSIONS: The ensemble model combined two different classification models for pneumonia that performed at 0.983 AUC for an external test dataset from a completely different data source. Furthermore, AI probability scores showed significant changes between cases of different clinical prognosis, which suggest the possibility of increased efficiency and performance of the CXR reading at the diagnosis and follow-up evaluation for pneumonia.