Li Yanxin, Shirley Ben C, Wilkins Ruth C, Norton Farrah, Knoll Joan H M, Rogan Peter K
CytoGnomix, Inc., POB 27052, 60 N. Centre Rd, London, ON, Canada.
Health Canada, Environmental and Radiation and Health Sciences Directorate, Ottawa, ON, Canada.
Radiat Prot Dosimetry. 2019 Dec 31;186(1):42-47. doi: 10.1093/rpd/ncy282.
Accuracy of the automated dicentric chromosome (DC) assay relies on metaphase image selection. This study validates a software framework to find the best image selection models that mitigate inter-sample variability. Evaluation methods to determine model quality include the Poisson goodness-of-fit of DC distributions for each sample, residuals after calibration curve fitting and leave-one-out dose estimation errors. The process iteratively searches a pool of selection model candidates by modifying statistical and filter cut-offs to rank the best candidates according to their respective evaluation scores. Evaluation scores minimize the sum of squared errors relative to the actual radiation dose of the calibration samples. For one laboratory, the minimum score for the curve fit residual method was 0.0475 Gy2, compared to 1.1975 Gy2 without image selection. Application of optimal selection models using samples of unknown exposure produced estimated doses within 0.5 Gy of physical dose. Model optimization standardizes image selection among samples and provides relief from manual DC scoring, improving accuracy and consistency of dose estimation.
自动双着丝粒染色体（DC）分析的准确性依赖于中期图像选择。本研究验证了一个软件框架，以找到减轻样本间变异性的最佳图像选择模型。确定模型质量的评估方法包括每个样本的DC分布的泊松拟合优度、校准曲线拟合后的残差以及留一法剂量估计误差。该过程通过修改统计和过滤截止值，迭代地搜索一组选择模型候选者，根据它们各自的评估分数对最佳候选者进行排名。评估分数使相对于校准样本实际辐射剂量的平方误差总和最小化。对于一个实验室，曲线拟合残差法的最低分数为0.0475 Gy²，而不进行图像选择时为1.1975 Gy²。使用未知暴露样本应用最佳选择模型产生的估计剂量在物理剂量的0.5 Gy范围内。模型优化使样本间的图像选择标准化，并减轻了手动DC评分的负担，提高了剂量估计的准确性和一致性。
