PROJECT SUMMARY/ABSTRACT Half of the man-made radiation exposure to the U.S. population can be attributed to CT. Thus, stakeholders have invested heavily in the reduction of CT doses. Recently, CT Deep Learning Reconstruction or Denoising (DLRD) techniques have become available that claim to enable dose reduction in CT, just as CT iterative reconstruction (IR) claimed before it. Like DLRD, IR showed dramatic apparent increases in image quality by reducing image noise. Subsequent clinical studies found that true dose reductions capabilities for IR were only 20-25% in soft tissues and organs. Greater dose reductions resulted in compromised signal detectability for low-contrast lesions (e.g., liver metastases). Now, DLRD algorithms are being used in clinical practice and appear to yield image quality superior to IR. But again, we are finding that excessive radiation dose reduction can introduce new artifacts and compromise lesion detectability. Our overall objective is to develop and validate methods that can quantitatively determine CT protocols that deliver the needed diagnostic performance at the lowest patient dose for any scanner model or reconstruction algorithm. In our first two competitive award periods, we developed robust Channelized Hotelling model observers (MOs) to quantify diagnostic performance using low contrast objects within a uniform phantom and validated the work with large scale reader studies for both filtered back projection and IR images. We are deploying these tools under EB028936 to allow robust dose optimization by CT users. However, there remain challenges. First, low-contrast phantoms used with most MOs are too simple. Second, highly realistic lesion and noise insertion tools require use of proprietary projection data and access to manufacturer software tools. Third, MOs must be extended to work with DLRD methods and be generalizable to any scanner. The proposed work will use patient image data, obviating the need for projection data, use our developed deep learning (DL)-MOs to achieve a scalable solution for performance assessment, and be generalizable to any algorithm or scanner. The goal of this renewal application is to develop robust DL-MO tools to efficiently predict diagnostic performance for images created with DLRD methods. We will accomplish this through three specific aims: 1. Develop and validate disease-insertion and low-dose-simulation tools for CT DLRD techniques. 2. Develop and validate DL-MO tools to quantify mean diagnostic performance for DLRD methods. 3. Extend DL-MO methods to estimate performance variations across readers. This work is the first to develop and use DL-MOs as accurate and efficient surrogates of human readers to characterize task-specific performance of DLRD methods in CT using patient images. The resultant performance assessment engine will facilitate training / testing of DL-based noise reduction algorithms and optimizing CT protocols and doses. These significant capabilities will ensure that diagnostic performance and the lifesaving diagnostic information obtained from CT are not compromised in DLRD-processed images.

PROJECT NARRATIVE Missing a liver lesion, or other subtle pathology, can have life-or-death consequences for patients. However, techniques that change CT images in a non-linear fashion are being commercially deployed based on simplistic assessments, such as noise reduction, rather than on maintenance of diagnostic performance. This work will provide the CT community tools to assess diagnostic performance for new CT imaging methods as they become available, facilitating optimization of dose reduction and ensuring safe clinical deployment.