Deep learning-based gantry angulation reduces eye lens exposure in brain CT

Abstract

<h2>Abstract</h2><h3>Background</h3> To develop and evaluate a deep learning–based method for delineating scan areas on CT localizers to reduce eye lens exposure during brain CT. <h3>Methods</h3> In this retrospective study, 2,175 adult brain CT scans were collected from internal and external cohorts. Ground-truth scan areas aligned to the supraorbitomeatal line were annotated in consensus by a radiographer and a radiologist. YOLOv11 rotated object detection models (N,S,M,L,X) were trained with five-fold cross-validation. Performance was evaluated using absolute angular difference (AAD) and the Dice similarity index. The best-performing model was applied to independent test cohorts using an ensemble. Predicted scan areas were used to reconstruct CT volumes, which were assessed for eye lens inclusion and cranial coverage and compared against clinical routine acquisitions. Statistical significance was tested with one-sided test for paired proportions, evaluating superiority in reducing eye lens inclusion and non-inferiority with respect to cranial coverage. <h3>Results</h3> During cross-validation, the YOLOv11-M model achieved near-perfect alignment with the ground truth (AAD, 1.11°±1.03°; Dice-index, 0.974±0.012). On independent test cohorts, predicted scan areas maintained high accuracy (AAD, 1.12–1.55°; Dice-index, 0.959–0.972). Reconstructed CT volumes demonstrated significant reductions in eye lens inclusion by 76.5% in the internal cohort and by 95.4% and 87.7% in the external cohorts. Cranial coverage was preserved, with rates comparable to routine practice, indicating non-inferior brain coverage. <h3>Conclusion</h3> Automated delineation of brain CT localizers using a deep learning model substantially reduced unnecessary eye lens exposure while maintaining cranial coverage. This approach may improve patient safety when integrated into clinical workflows.

Ähnliche Arbeiten