From patient-informed to patient-specific organ dose estimation in clinical computed tomography

Many hospitals keep a record of dose after each patient's CT scan to monitor and manage radiation risks. To facilitate risk management, it is essential to use the most relevant metric, which is the patient-specific organ dose. The purpose of this study was to develop and validate a patient-specific and automated organ dose estimation framework. This framework includes both patient and radiation exposure modeling. From patient CT images, major organs were automatically segmented using Convolutional Neural Networks (CNNs). Smaller organs and structures that were not otherwise segmented were automatically filled in by deforming a matched XCAT phantom from an existing library of models. The organ doses were then estimated using a validated Monte Carlo (PENELOPE) simulation. The segmentation and deformation components of the framework were validated independently. The segmentation methods were trained and validated using 50-patient CT datasets that were manually delineated. The deformation methods were validated using a leave-one-out technique across 50 existing XCAT phantoms that were deformed to create a patient-specific XCAT for each of 50 targets. Both components were evaluated in terms of dice similarity coefficients (DSC) and organ dose. For dose comparisons, a clinical chest-abdomen-pelvis protocol was simulated under fixed tube current (mA). The organ doses were estimated by a validated Monte Carlo package and compared between automated and manual segmentation and between patient-specific XCAT phantoms and their corresponding XCAT targets. Organ dose for phantoms from automated vs. manual segmentation showed a (similar to)2% difference, and organ dose for phantoms deformed by the study vs. their targets showed a variation of (similar to)5% for most organs. These results demonstrate the great potential to assess organ doses in a highly patient-specific manner.