Internal feasibility evaluation of SAM-Med3D for cervical cancer clinical target volume segmentation on planning computed tomography.

Abstract

Accurate delineation of cervical cancer clinical target volume (CTV) remains labor-intensive and variable. This retrospective single-center study internally evaluated a SAM-Med3D baseline initialized from public pretrained weights and fine-tuned on the study training cohort for prompted cervical cancer CTV segmentation on planning CT. The eligible cohort comprised 182 cases split into training ([Formula: see text]), validation ([Formula: see text]), and temporally subsequent independent test ([Formula: see text]) sets. CT scans were resampled, normalized, and prepared as label-centered 128 × 128 × 128 patches. The SAM-Med3D model was initialized from public pretrained weights and fine-tuned on the training cohort without architectural modification, then assessed with simulated/oracle-guided prompts at 1, 3, 5, 7, 9, and 11 clicks using Dice, HD95, 3-mm surface Dice, and volume consistency. Mean Dice increased from 0.827 at 1 click to a peak of 0.835 at 7 clicks and was 0.833 at 11 clicks. HD95 decreased from 12.46 mm at 1 click to 9.48 mm at 9 clicks and 10.18 mm at 11 clicks, and 3-mm surface Dice increased from 0.707 to 0.721 at 7 clicks. Thus, multi-click prompting produced modest improvements that plateaued at later clicks. A supplementary risk-guided prompt-selection pilot was performed on the same test set but was treated only as exploratory. Because patch extraction and prompt generation used reference-contour information, all analyses represent controlled upper-bound evidence. This study provides an internal feasibility baseline for promptable cervical cancer CTV segmentation; label-free ROI selection, full-volume inference, external validation, and clinician-in-the-loop testing remain necessary before clinical deployment.