Program Information
Building An in Silico Imaging Trial That Replicates An Existing Human Trial
A Badano*, A Badal , S Glick , C Graff , F Samuelson , D Sharma , R Zeng , U.S. Food & Drug Administration (CDRH/OSEL), Silver Spring, MD
Presentations
SU-F-702-2 (Sunday, July 30, 2017) 2:05 PM - 3:00 PM Room: 702
Purpose: To describe the design of a virtual imaging clinical trial for regulatory evaluation with the goal of replicating results of an existing human trial for evaluating safety and effectiveness of breast tomosynthesis as compared to digital mammography.
Methods: Computational models for imaging are not new but are seldom used in regulatory evaluations. On one hand, human clinical trials approximate real-world device use by relying on patient data and device performance but suffer from limited sample sizes, significant variability, and challenges in truth determination with added patient radiation risk. On the other hand, in silico trials suffer from a somewhat simplified version of all components but benefit from unlimited data, controllable component variability, exact knowledge of the truth, and no additional risk for patients making them least burdensome. Replication of an existing human clinical imaging trial requires the design of the following components: distribution of patients and pathology, system modeling, design of readers and their variability, and statistical analysis of comparative results.
Results: In project VICTRE, we selected a patient population distribution to match the screening population used in the human study with compressed breast thickness from 3.5 to 6.0 cm and a range of apparent breast density corresponding to the 4 BIRADS categories. Cancer lesions were represented by a spiculated mass and a microcalcification cluster designed to match pathology found in the human trial adjusted to control absolute performance of the FFDM image interpretation algorithms. Uncertainty was estimated using fully-paired MRMC analysis with computational readers trained on distinct image sets. The study was sized to achieve uncertainties lower than those seen in the human trial for the endpoint, the differential performance in terms of area-under-the-ROC curve.
Conclusion: This work seeks to stimulate discussion on how to increase the use of computational modeling in imaging system evaluation for regulatory purposes.
Contact Email: