Comparison of ROC methods for partially paired data

Brandon D. Gallas; Lorenzo L. Pesce

doi:10.1117/12.813688

Translator Disclaimer

12 March 2009 Comparison of ROC methods for partially paired data

Brandon D. Gallas, Lorenzo L. Pesce

Author Affiliations +

Proceedings Volume 7263, Medical Imaging 2009: Image Perception, Observer Performance, and Technology Assessment; 72630V (2009) https://doi.org/10.1117/12.813688
Event: SPIE Medical Imaging, 2009, Lake Buena Vista (Orlando Area), Florida, United States

Abstract

In this work we investigate ROC methods that compare the difference in AUCs (area under the ROC curve) from two modalities given partially paired data. Such methods are needed to accommodate the real world situations, where every case cannot be imaged or interpreted using both modalities. We compare variance estimation of the bivariate binormal-model based method ROCKIT of Metz et al., as well as several different non-parametric methods, including the bootstrap and U-statistics. This comparison explores different ROC curves, study designs (pairing structure of the data), sample sizes, case mix, and modality effect sizes.

Citation Download Citation

Brandon D. Gallas and Lorenzo L. Pesce "Comparison of ROC methods for partially paired data", Proc. SPIE 7263, Medical Imaging 2009: Image Perception, Observer Performance, and Technology Assessment, 72630V (12 March 2009); https://doi.org/10.1117/12.813688

ACCESS THE FULL ARTICLE