A neural network-enhanced methodology for the rapid establishment of local DRLs in interventional radiology: EVAR as a case example.

Purpose: To develop a neural network-enhanced workflow for the automatic and rapid establishment/update of local diagnostic reference levels (DRLs) in interventional radiology (IR) using endovascular aneurysm repair (EVAR) procedures as a case example.

Methods: Radiation dose reports were collected retrospectively for 46 consecutive EVAR procedures. These reports served as demonstrative data for the development of the proposed methodology. An algorithm was developed to receive multiple dose reports, automatically extract the kerma area product (KAP), air kerma (Ka,r), number of exposure images, and fluoroscopy time (FT) from each report and calculate the first, second, third quartiles as well as the maximum and minimum values of the extracted parameters. To extract the values of interest from the dose reports, Tesseract, an open-source optical character recognition (OCR) engine was employed. Furthermore, the accuracy and time efficiency of the proposed methodology were assessed. Specifically, the values extracted from the algorithm were compared with the ground truth values and the algorithm's processing time was compared with the respective time needed to manually extract and process the values of interest.

Results: The OCR-based algorithm managed to correctly recognize 182 from the 184 target values, resulting in an accuracy of 99%. Moreover, the proposed pipeline reduced the processing time for the establishment of DRLs by 98%. DRL value for EVAR procedures, set as the third quartile of KAP was found to be 551 Gy*cm.

Conclusion: An accurate and time-efficient workflow was developed for the establishment of local DRLs in interventional radiology.