| Abstract Scope |
Cavities are a significant defect in irradiated materials. They have been extensively studied across various material systems using different imaging techniques to assess swelling and inform models. The characterization and quantification of cavities is influenced by several factors, including imaging conditions, cavity density, and cavity size, which inherently involve measurement error and uncertainty. Numerous studies have suggested that there is significant variability in cavity detection and that it arises largely from differences in how individual labelers—varying by their training—identify cavities. Traditionally, this variability has been ignored when comparing results. Here, we first identify the source and scale of labeler-based uncertainties. Second, we propose an innovative solution to minimize labeler variability through non-expert crowdsourcing via Amazon Mechanical Turk (MTurk) workers and a machine learning model (YOLOv11m) trained on synthetically generated bounding boxes for final label aggregation and ground truth formation. Our proposed method achieves F1 scores – a metric for precision and accuracy in object detection – comparable to those achieved by expert labelers, thus demonstrating improved cavity detection. Moreover, the method excels in addressing the detection of ambiguous cavities—those not unanimously agreed upon by experts. Third, through large-scale sampling and simulated TEM images, we have systematically identified that background contrast significantly affects the identification of these ambiguous cavities, increasing the uncertainty in images with heightened background contrast. We conclude by proposing a method for flagging images with probable high uncertainties through a dynamic threshold method to improve overall quantified cavity databases when the proposed crowdsourcing method is not feasible. |