Quality Evaluation Method for Chest X-Ray Images using the Reference Patterns

Abstract

Chest X-Ray image quality evaluation relies on clinicians' judgment with clinical guidelines and expertise. Recently, the deep learning model for disease diagnosis has been made using X-ray image datasets. In this case, the quality of the image, which is the training model's input, determines the diagnostic model's performance. However, there are cost and time limitations to relying solely on the evaluation of clinicians for the massive amount of images. Therefore, there is a need to automatically evaluate the checklist of a given guideline and select a high-quality image. This paper generates a high-quality pattern from the chest X-ray image, and the difference between pixels from the target image is imaged. It is a method of evaluating quality with a CNN-based model by patterning the pixel value distribution of the generated image. The proposed method shows that it is possible to evaluate multiple integrated items, not a single evaluation item. In addition, to confirm the effectiveness of the proposed model, data composed by refining only randomly configured data and high-quality data without applying quality evaluation were applied to the Nodule Diagnostics Model. As a result of the experiment, model performance improved by 19% accuracy and 21% AUC when training with high-quality data.