Room: AAPM ePoster Library
Purpose: To better understand the relationship between Hounsfield Unit (HU) and electron density (ED) in computed tomography (CT) imaging. By investigating the biology, chemistry, and physics underlying the degeneracy seen in the HU to ED conversion process, we hypothesize that ED accuracy can be significantly improved by introducing additional anatomic/biologically based information.
Methods: The vast majority of molecular constituents in tissues are water, carbohydrates, lipids, proteins, and minerals/hydroxyapatite with the exact proportions being tissue/organ dependent. In this work, we modeled the radiological properties of these molecules and human tissues by calculating the mass density normalized HU, HU?, and mass density normalized ED, ED?, for each at kV and MV energies. HU? vs ED? for kVCT was found to be highly degenerate but not when considering a multi-dimensional relationship between HU, molecular/anatomic composition, and ED. MVCT was linear and did not exhibit an additional molecular/anatomic dependency for HU? vs ED?.
Results: CycleGAN machine learning was used to implicitly model this multi-dimensional relationship between HU, molecular/anatomic composition, and ED for kV SECT by learning the kVCT<->MVCT/ED relationship. 120 head and neck cancer (H&N) patients with kVCT and MVCT scans were retrospectively recruited for this study with 100 patients used for model training alone. This model was then applied to the remaining 20 patient datasets/scans for testing/validation. Prior to machine learning, mean differences/errors (ME) of kVCT determined ED versus MVCT determined ED were 2% and 8% for soft tissue and bone, respectively. Post-machine learning, ME of kVCT determined ED were <0.5% for all tissue types.
Conclusion: The kVCT HU versus ED relationship is degenerate until molecular/tissue composition is considered. The implicit molecular/tissue information learned through machine learning can be used to improve ED calculation in kVCT to accuracy better than 0.5%. Our approach may be helpful in reducing range uncertainty for particle therapy.
Funding Support, Disclosures, and Conflict of Interest: Research reported in this abstract was supported by the NIBIB of the National Institutes of Health under award number R21EB026086.
Not Applicable / None Entered.