Establishing discard criteria for lead aprons using deep learning-based quantification of defect area on X-ray fluoroscopic video
摘要
To establish an objective discard framework for lead aprons, defect areas quantified from X-ray fluoroscopy videos were linked to their corresponding dosimetric impacts. A fine-tuned YOLOv8 instance segmentation model, achieving an average precision at an intersection-over-union threshold of 0.5 (AP@0.5) of 0.457, was employed to analyze fluoroscopic videos of 18 aprons with 0.25 mmPb equivalence. The model enabled automated defect detection and area estimation with a video processing time of 15 s. Dosimetry experiments mimicking endoscopic retrograde cholangiopancreatography (ERCP) conditions were conducted using an anthropomorphic phantom and apron samples with slit defects of varying widths. The results indicated that the transmitted dose increased linearly with defect area under all tested conditions. Based on weighted least-squares analysis of covariance (WLS-ANCOVA), slit widths of 1.0, 2.0, and 3.5 mm were pooled for modeling, as they exhibited similar dosimetric behavior. A linear model was applied to relate defect area to the increase in transmitted dose, defining the discard threshold as the minimum area at which the upper one-sided 95% prediction limit reached a specified dose limit. Given a baseline dose of 2.38 mSv after transmission through the apron without defects and a regulatory dose limit of 5 mSv per 3 months, the discard threshold for slit-like defects was determined to be 2.87 cm2. The proposed framework enables automated and facility-tailorable discard decisions by linking quantified defects to radiation protection objectives using conservative, prediction-based criteria.