[Automation of Damage Detection and Damage Area Measurement of X-ray Protective Clothing Using Deep Learning]

Damage to shielding sheets on X-ray protective clothing may be a cause of increased radiation exposure. To prevent increased radiation exposure, periodic quality control of shielding sheets is needed. For quality management, a record of the size of damage is required after checking for the existence of damage, and this requires a great deal of effort and time. Additionally, the detection model created from the images of the shielding sheets, limited by the number of samples, is predicted to have a low detection precision. The purpose of this study was to automate damage area detection and area measurement using artificial damage images and a damage detection model created using deep learning.

By synthesizing the X-ray protective clothing CT localizer image and the image simulating damage, we created an artificial damage image. We then found the detection precision of the damage detection model created by the artificial damage image and YOLOv5s, and error of the automatically measured damage area.

The accuracy rate of the damage detection model was 0.746, the precision was 0.645, the reproduction rate was 0.741, the F value was 0.690, and 48 mm² or the detectable area of damage ranged from 2 mm² to 113 mm². The mean value of the damage area error was 7.58% for areas not including the hem and 43.39% for areas including the hem. In the areas not including the hem, with a detected damage area of 91%, the damage area error was 0%. Additionally, the process from damage area detection to damage area measurement was completed in 20 seconds.

By using a damage detection model created with only artificial damage areas, it was possible to automate damage detection and damage area measurement, and this saved time for X-ray protective clothing management.