TOPOS: target organ prediction on scout views for automated CT scan planning.
Authors
Affiliations (4)
Affiliations (4)
- Department of Diagnostic and Interventional Radiology, Klinikum rechts der Isar, Technical University of Munich, School of Medicine and Health, TUM University Hospital, Munich, Germany. [email protected].
- Department of Diagnostic and Interventional Radiology, Klinikum rechts der Isar, Technical University of Munich, School of Medicine and Health, TUM University Hospital, Munich, Germany. [email protected].
- Department of Diagnostic and Interventional Radiology, Klinikum rechts der Isar, Technical University of Munich, School of Medicine and Health, TUM University Hospital, Munich, Germany.
- Department of Cardiovascular Radiology and Nuclear Medicine, Technical University of Munich, School of Medicine and Health, German Heart Center, TUM University Hospital, Munich, Germany.
Abstract
Overscanning is a common issue in CT planning, leading to unnecessary radiation exposure. To develop a deep learning model to segment anatomical structures in scout views to optimize scan ranges and reduce radiation. In this single-center retrospective study, 1146 patients undergoing CT between 2022 and 2025 were included. The model was trained on segmentations of 26 target structures in five regions (head, neck, chest, chest-to-pelvis, abdomen-to-pelvis), transferred to scout views and manually corrected. Performance was evaluated using the Dice-Sørensen coefficient and normalized surface distance on an internal test set of 100 patients and 36 external chest CTs. Automated versus manual scan planning was compared in 61 internal (chest, upper abdomen, head) and 14 external (chest) CTs, with z-axis coverage and dose-length product. 1146 patients (mean age, 63 ± 17 years; 577 men) were included. For target structures in five regions, mean DSC and NSD were 0.93 and 0.88. External mean DSC across chest targets was 0.851 ± 0.051. Automated planning captured relevant anatomy in 98% of internal and 92.9% of external CTs. Scan length significantly decreased for automated planning in the internal test cohort (chest 50 mm (15%), p < 0.001; upper abdomen 60 mm (25%), p < 0.001; cranial 19 mm (11%), p < 0.001), yielding corresponding DLP reductions of 19%, 25% and 11%, respectively. In the external cohort, scan length decreased by 115 mm (28.7%, p < 0.0001) with a corresponding 28.5% DLP reduction. The proposed model enables reliable automated CT scan planning and reduces overscanning and radiation exposure without compromising diagnostic quality. Question Can segmentation of anatomical structures on CT scout views enable automated scan planning to reduce overscanning and unnecessary radiation exposure? Findings A deep learning model segmented planning-relevant anatomy on CT scout views and reduced scan length and dose while preserving anatomical coverage. Clinical relevance Unnecessary radiation from overscanning is a patient safety concern. Automated CT planning with the proposed deep learning model reduces radiation exposure while ensuring full anatomical coverage for the evaluated scan regions.