[Measurement of Defect Areas in Lead Aprons Using Click-prompt-based Semi-automatic Region Extraction with the Segment Anything Model].

PURPOSE: To develop a click-prompt-style semi-automatic defect segmentation method using Segment Anything Model 2 (SAM2) with Low-Rank Adaptation (LoRA) for fluoroscopic images of lead aprons and to evaluate its quantitative performance and usefulness.

METHODS

Twenty-four fluoroscopic images acquired from 19 lead aprons were divided into training/validation (n=14) and test datasets (n=10). Two radiological technologists independently annotated the test images and established Consensus GT masks after reviewing discrepancies. Four radiological technologists performed manual and click-prompt-based segmentation using SAM2 Original and SAM2-LoRA (r=8, 16, and 32). All methods were evaluated against the Consensus GT.

RESULTS

Inter-annotator agreement between the two independent GT masks was ICC(2,1)=0.982 for area, with a Dice coefficient of 0.802 and IoU of 0.674. Mean Dice coefficients against the Consensus GT were 0.708 for manual segmentation, 0.426 for SAM2 Original, and 0.752, 0.774, and 0.759 for SAM2-LoRA with r=8, 16, and 32, respectively. The r=16 setting showed the smallest mean absolute area error among the models (69.7 mm2).

CONCLUSION

Under the conditions of this small single-center study, click-prompt-based SAM2-LoRA may support measurement of candidate defect areas in fluoroscopic images of lead aprons. External validation is required before use for automated rejection decisions.