A Dynamic Patient Data Library and Virtual Imaging Trial Platform for Training and Evaluating AI-based Algorithms in CT Imaging - PROJECT SUMMARY/ABSTRACT Deep Learning Reconstruction and Noise Reduction (DLR) algorithms are increasingly used in CT clinical practice, with substantial potential to reduce radiation dose and improve image quality. However, their advancement is hindered by the lack of accurate diagnostic performance evaluation tools and limited public CT projection data for training. These challenges create significant hurdles for developers to optimize their algorithms and prove dose reductions, and for regulatory bodies and clinical users to validate these claims, raising crucial safety concerns for millions of CT patients. With prior funding support from the NIBIB, we created a CT patient projection data library in an open, vendor-neutral format, DICOM-CTPD. This resource has been globally utilized by researchers to develop, train, validate, and test various DLR algorithms, including those mark the beginning of AI applications in CT imaging. Despite its great success, this library is constrained by its static dose levels, disease conditions, and data from older scanner models. These limitations highlight the pressing need for a new, dynamic patient projection data library with advanced diagnostic performance evaluation tools. The objective of this project is to develop a dynamic, customizable, vendor-neutral patient projection data library and software platform (DICOM- CTPD-VIT). This platform will be capable of generating a diverse range of conditions in radiation dose, reconstruction parameters, and lesion characteristics, and will facilitate automated image quality and diagnostic performance evaluation. The project comprises three specific aims: Aim 1: Develop a dynamic and vendor-neutral patient projection-data library with an integrated virtual- imaging-trial (VIT) software platform (DICOM-CTPD-VIT). Aim 2: Develop and validate a patient-data-based VIT evaluation framework for DLR methods. Aim 3: Develop and validate VIT-based approaches to quantifying hallucinations in DLR methods. The innovation and significance of this project lie in its dynamic, vendor-neutral patient projection data library, integrated with a VIT, and a patient-data-based evaluation framework. These resources allow for creation of varied imaging conditions, tailored to the needs of clinical users and researchers. It provides pathways for both algorithm development and clinical evaluation, addressing the diverse needs within the CT field to facilitate effective development and safe implementation of DLR methods in clinical practice. Moreover, the platform has the potential to generate an unlimited number of cases with ground truth, useful for training and evaluating other AI-based diagnostic tools beyond DLR.
