Author: Maria Jose Medrano, Grant Stevens, Liyan Sun, Justin Ruey Tse, Adam S. Wang, Sen Wang 👨🔬
Affiliation: Department of Radiology, Stanford University, GE HealthCare, Stanford University 🌍
Purpose: Patient exposure to ionizing radiation is a major concern in CT imaging. Size-specific dose estimation methods can prospectively estimate organ-level radiation doses based on patient sizes and tube current modulation (TCM). However, their accuracy depends on organ segmentations that are not prospectively available. This study introduces an updated Scout-Net model for direct prospective estimation of organ-level doses for any TCM map and compares its performance to two previously established methods.
Methods: We analyzed CT scans of 130 adult patients undergoing contrast-enhanced chest-abdomen-pelvis examinations on a GE Revolution scanner with 120 kVp and TCM. Reference organ doses for six organs (lungs, kidneys, liver, pancreas, bladder, spleen) were retrospectively calculated using open-source tools MC-GPU and TotalSegmentator. Our updated Scout-Net model predicted organ-level doses for any TCM profile by leveraging the discrete cosine basis representation of TCM maps and using lateral and frontal scouts and patient scan range as inputs. Two established methods were implemented for comparison: size-specific dose estimation following AAPM Task Group 204 (Global CTDIvol) and its version adapted for TCM and organ-specific estimates (Organ CTDIvol). A 5-fold cross-validation assessed the generalizability of all three methods by comparing mean absolute percentage dose errors across the six studied organs.
Results: Overall absolute percentage errors for the six organs were 15.9% (Global CTDIvol), 8.8% (Organ CTDIvol), and 7.1% (Scout-Net). The largest discrepancies were observed in bladder dose errors: 27.0%, 13.4%, and 9%, respectively. Scout-Net consistently outperformed both established methods across all organs, except for lung and kidney dose estimation, where it showed comparable performance to Organ CTDIvol.
Conclusion: Our updated Scout-Net model demonstrated superior accuracy compared to Global CTDIvol and comparable to better performance relative to Organ CTDIvol, without requiring organ segmentations during inference. These results highlight Scout-Net’s potential as a clinically useful tool for prospective organ-level dose estimation in CT imaging.