Automated Deep Learning-Based Patient Positioning in Chest CT Using AP Localizer Images

Overview

This dual-center study developed and validated a deep learning (DL) approach to automatically detect patient body centerline displacement from the gantry isocenter using only anterior-posterior (AP) localizer images in chest CT. The DL models demonstrated accurate patient centering, potentially reducing radiation dose and improving image quality compared to manual or camera-based methods.

Background

Computed tomography (CT) requires precise patient positioning to optimize image quality and minimize radiation exposure. Mis-centering of patients along the table height (Y-axis) is common and can increase radiation dose by up to 30% and degrade image quality. Existing automatic positioning methods often rely on 3D cameras, which have limitations including calibration complexity and errors caused by external objects on patients. Deep learning has shown promise in automating various CT tasks but has been scarcely applied to patient positioning.

Data Highlights

Key Findings

• Patient mis-centering in chest CT is highly prevalent, with prior studies reporting average errors up to 43 mm and dose increases up to 30% due to mis-centering.
• Deep learning models trained separately on data from two different CT scanners (Siemens and Phillips) using only AP localizer images achieved accurate detection of patient centerline displacement.
• The study included 5754 chest CT cases after excluding images with truncation artifacts, ensuring reliable input data for model training and testing.
• Compared to camera-based automatic positioning methods, the DL approach avoids calibration challenges and errors caused by external objects on patients.
• Automated positioning using DL has the potential to improve image quality by reducing noise and enhancing lesion signal-to-noise ratio while decreasing radiation dose.

Clinical Implications

Implementing deep learning-based automatic patient positioning using only AP localizer images can streamline CT workflow by eliminating the need for additional hardware like 3D cameras. This approach may reduce patient mis-centering errors, thereby lowering radiation dose and improving diagnostic image quality. It offers a practical solution adaptable to different CT scanners without complex calibration.

Conclusion

The study demonstrates that deep learning applied to AP localizer images can accurately automate patient positioning in chest CT, addressing a critical source of dose inefficiency and image degradation. This method holds promise for widespread clinical adoption to enhance CT safety and efficacy.

References

1. Akintayo et al 2021 -- Prevalence of Patient Mis-centering in CT Scans
2. Sukupova et al 2020 -- Impact of Patient Mis-centering on CT Dose
3. Li et al 2019 -- Dose Increment Due to Patient Mis-centering
4. Furukawa et al 2018 -- Effect of Table Height on Tube Current Modulation
5. Euler et al 2017 -- Organ Dose Changes from Table Mis-centering
6. Kaasalainen et al 2016 -- Organ Dose Variation with Patient Positioning
7. Booij et al 2015 -- Automatic Patient Positioning Using 3D Camera
8. Dane et al 2014 -- Patient Positioning Accuracy with 3D Camera and AP Localizer
9. Gang et al 2013 -- Impact of Mis-centering Correction on Image Quality and Dose
10. Saltybaeva et al 2012 -- Reduction of Positioning Errors Using 3D Camera
11. Deep Learning in Medical Imaging -- Various Authors