Motivation Computer aided triaging (CADt) devices for intracranial hemorrhage (ICH) in the emergency room (e.g. Rapid ICH K221456) is one important example where pediatric and adult cases exist in a reading queue where pediatric patients could be disadvantaged by being deprioritized for time sensitive treatment using an adult-trained AI model that extrapolates poorly to pediatric patients. While these AI/ML devices have potential to benefit pediatric patients, there is currently a lack of annotated pediatric data for evaluating the balance of risk and benefits.

Purpose To address data availability challenges, we propose to supplement available pediatric patient computed tomography (CT) datasets with data generated in silico, generated using realistic computational human models and physics-based CT simulations. In silico data generation allows for creating examples with true labels with a fraction of the cost that is needed to label real patient data.

We have previously combined the pediatric and adult digital XCAT cohort of phantoms with the XCIST x-ray CT simulation framework to create realistic CT exams. This preliminary work was in support of investigating the effectiveness of deep learning denoising algorithms in pediatric patients. Dr. Elena Sizikova also built a pipeline for comparative evaluation of digital mammography AI using in digital models of the breast and digital mammography (DM) acquisition devices, which will serve as a starting point for this aim. Specifically, we will rely on the XCAT phantom as a digital model of the pediatric brain. We will rely on the XCIST simulator to generate CT images. Specifically, we will vary the following parameters:

Digital model: patient size, age, contrast b/w grey and white matter, skull hardness, thickness, and ICH morphology, texture, and location

Imaging Parameters (CT): Radiation dose (MA, KV voltage), slice thickness, reconstruction kernels, reconstruction field of view (FoV).

pip install -r notebooks/requirements.txt

Tested on python 3.11.3

• REALYSM_PedCT: pedsilico-pilot.ipynb: CT simulation pipeline that we aim to build off of for this project, in particular this notebook was used to make the pilot data images shown in Methods)
• pediatricIQphantoms: running_simulations.ipynb: examples of using a Python wrapper around the Michigan Image Reconstruction Toolbox (MIRT) for simple, faster CT simulations