Personalized eye protection for head CT organ-based tube current modulation: A deep learning approach to derive 3D eyeball models from a single-view topogram.

BACKGROUND: The lens of the eye is highly radiosensitive, yet personalized shielding during head CT remains challenging due to the lack of a rapid, pre-scan localization method. PURPOSE: To develop and validate a deep learning solution that enables automated, patient-specific eye protection by generating a precise 3D eyeball model directly from a single-view topogram. METHODS: Our two-stage approach combines an advanced data simulation pipeline-which generates realistic training topograms from digitally reconstructed radiographs (DRRs) using a table-movement-aware model and CycleGAN-based stylization-with a dedicated generative network (EyeGen-Net). The model was trained on 400 synthetic and validated on 100 real clinical samples. RESULTS: EyeGen-Net achieved a Dice Similarity Coefficient of 0.79 ± 0.08, a Hausdorff Distance of 5.40 ± 1.57 mm, and an Average Surface Distance of 1.84 ± 0.65 mm against expert segmentations. Crucially, phantom validation demonstrated that the derived 3D model facilitates organ-based tube current modulation (OBTCM), yielding an approximate 30% reduction in lens dose across different scanning modes without compromising diagnostic image quality. CONCLUSIONS: This work provides a practical, automated pathway for implementing personalized radioprotection in routine head CT, aligning with the ALARA (As Low As Reasonably Achievable) principle.