Predicting elevated natriuretic peptide in chest radiography: emerging utilization gap for artificial intelligence

Figures

• 

  Figure 1. Chest X-ray images and featured map of weak learners. (A) Chest X-ray image with a BNP of 9 pg/mL and a GRAD-CAM image generated by the EfficientNetV2L-based model. (B) The EfficientNetV2L-based-model identified features such as pulmonary congestion and pacemaker, with a BNP of 798 pg/mL. (C) The attention map revealed the Vit-b16-based-model’s features in the caption, the pulmonary vessel, and the diaphragmatic line, with a BNP of 413 pg/mL. The caption in the upper-left corner is a Kanji character meaning the supine position. (D) The EfficientNetV2S-based-model identified features in the captions of images with a BNP of 824 pg/mL. The caption in the upper-left corner is a Kanji character meaning the sitting position. (E) The other EfficientnetV2S-based model did not focus on the features in the captions but on the pleural effusion, cardiomegaly, pulmonary artery, air bronchogram, and Kerley A lines. The caption in the upper-left corner is a Kanji character meaning the sitting position.

• 

  Figure 2. Performance of the models. The ROC curves (A) and PR curves (B) of the AI models for predicting BNP ≥ 200 pg/mL are shown. Similarly, the ROC curves (C) and PR curves (D) for predicting BNP ≥ 100 pg/mL are shown. The ROC and PR curves were shown for the final ensemble model as well as for 3 of the 31 weak leaners.

• 

  Figure 3. Representative images and predictions of the AI models and humans (A, B, C, D, and E). The chest X-ray images and their featured map images, age, sex, BNP value, AI prediction, and human accuracy are shown. The featured map images were generated by the EfficientNetV2S-based model using GRAD-CAM. The captions in the upper-right corner of (A) and the upper-left corner of (B, C, and D) are Kanji characters meaning sitting and standing positions, respectively.

• 

  Figure 4. The results of the tests. (A and B). Both the early-career medical professionals (0.692 ± 0.042 vs. 0.851 ± 0.074, P < 0.001) and veterans in the medical career (0.728 ± 0.051 vs. 0.803 ± 0.054, P < 0.001) improved their accuracy with the assistance of the AI model. (C) The increase in the accuracy with the AI assistance was greater for the early-career professionals than for the veterans (0.159 ± 0.069 vs. 0.074 ± 0.052, P < 0.001). (D) In the non-assisted test, the accuracy had a weak positive correlation with the duration of medical careers (r = 0.414, P = 0.014). (E) However, in the AI-assisted test, it had a weak negative correlation (r = −0.347, P = 0.041).