(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Optimization of medical imaging system performance should be guided by task-based measures of image quality, which quantify the ability of an observer to perform specific tasks. The estimation ROC curve (EROC) has been proposed for evaluating an observer on general joint detection/estimation tasks. However, the EROC ideal observer (IO) test statistic cannot be directly approximated using supervised learning approaches. In this study, we proposed a joint supervised learning and Markov-Chain Mote Carlo (MCMC) strategy to approximate EROC-IO.
Methods: The IO test statistic can be decomposed into a multiplication of ideal likelihood ratio and utility weighted posterior mean. A convolutional neural network (CNN) with two channels is constructed for approximating the ideal estimate and the ideal likelihood ratio. The ideal test statistic can be approximated by multiplying the ideal likelihood ratio and the utility weighted posterior mean, which can be estimated using the Monte Carlo integration on the ideal estimate.
Results: The ability of the proposed method is explored under background-known-exactly (BKE) and background-known-statistically (BKS) cases. Two different lumpy backgrounds were considered in BKS case. The EROC curves produced by the proposed method are compared to those produced by analytical computation when feasible. The area under the EROC curve (AEROC) is utilized to evaluate the observer performance.
The AEROC of the approximated IO under BKE case and BKS case with type 2 lumpy background are close to but still smaller than that of the analytical IO. The AEROC of the approximated IO under BKS case with type 1 lumpy background is much greater than that of channelized joint observer. These results indicate that the proposed strategy can successfully approximate IO test statistic and estimate.
Conclusion: The proposed method is an alternative approach to conventional numerical approaches and can approximate the ideal EROC observer test statistic and estimate in complex cases.
Funding Support, Disclosures, and Conflict of Interest: This research was supported in part by NIH awards EB020604, EB023045, NS102213, EB028652, R01CA233873, R21CA223799, and NSF award DMS1614305.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
