1194P - Development of a novel artificial intelligence (AI) algorithm to detect pulmonary nodules on chest radiography

Background

Artificial intelligence (AI) based on deep learning and convolutional neural networks (CNN) has been applied to various medical fields. We are developing a novel AI system to support the detection of lung cancer, which will enable physicians to efficiently interpret radiograms and perform diagnosis.

Methods

For training data, we analyzed the image features of 800 chest X-ray images (acquired mainly for lung cancer screening; 400 normal images and 400 abnormal images) from Fukushima Preservative Service Association of Health and 5 000 chest radiographs from the National Institutes of Health (NIH) database. We categorized these data into two groups. Group A included both 800 images and NIH database, and group B included only 800 images. Then, using ImageNet, we used these datasets to develop a proprietary AI algorithm with deep learning and a CNN, and analyzed the statistical accuracy of interpretations of the radiograms. We also demonstrated abnormal shadow in the form of a heat map display on each chest radiograph for easy visualization, and presented the positive probability score as an index value from 0.0 to 1.0 indicating the possibility of lung cancer. The accuracy of our AI system was improved by using technology to account for differences in radiographic apparatus and imaging environments.

Results

Using a positive probability cutoff value of 0.5, our novel AI showed an area under the curve (AUC) of 0.74, sensitivity of 0.75, and specificity of 0.60 for group A, and AUC of 0.79, sensitivity of 0.72, and specificity of 0.74 for group B. The accuracy on both groups was superior to that of radiologists (AUC 0.72), and was also compatible with previous study reports (AUC 0.67-0.78). The number of training data did not make a difference in accuracy. The heat map display was also clearly demonstrated on the monitor screen if a roentgenogram had abnormal shadows.

Conclusions

In this study, we confirmed that our proprietary AI system had a similar accuracy in the interpretation of chest radiographs to that of previous studies and radiologists. However, further research and improvement is needed to verify the accuracy. We are now in the process of performing various types of validation.

Clinical trial identification

Editorial acknowledgement

Legal entity responsible for the study

The authors.

Funding

Grants-in-Aid for Scientific Research in Japan.

Disclosure

All authors have declared no conflicts of interest.