Feasibility of AI-based Automated Identification of Sacral Anatomy and Lead Features During Sacral Neuromodulation Procedures.

Abstract

Objective: To evaluate the ability of artificial intelligence-based modeling to identify pertinent sacral anatomy and lead characteristics on fluoroscopic imaging.

Methods: An expert practitioner manually reviewed and segmented image pairs from sacral neuromodulation device placements between November 2017 and August 2022. These were used to train (n=177 lateral, 170 AP) and test (n=43 lateral, 36 AP) two semantic segmentation networks via nnUNet. A subset of this data was selected to train a model for electrode segmentation (n=201 train, n=38 test). Dice scores and subjective review were used to assess correlation between expert annotation and automated AI segmentation.

Results: The AI algorithm accurately segmented pertinent anatomic landmarks in 97% (35/36) of AP images and 93% (40/43) of lateral images based on subjective binary expert review, indicating a high level of clinically acceptable segmentations. When evaluated by 3-point Likert scale, the mean (SD) score was 2.86 (0.42) for AP images and 2.77 (0.57) for lateral images. The overall Dice score for all lateral landmarks was (mean ± std) 0.759 ± 0.067, with individual landmark scores ranging from 0.524 (S2) to 0.769 (S3 hillock). The overall Dice score (integrated) for all AP landmarks was 0.711 ± 0.117, with individual landmark scores ranging from 0.633 (medial edge of foramen) to 0.745 (midline of spine). The overall Dice score (integrated) for all electrodes in the combined AP and lateral imaging dataset was 0.889 ± 0.035.

Conclusions: Creating segmentation networks capable of producing high-quality annotations of the pertinent sacral anatomy and leads during sacral neuromodulation device placement is feasible. Further validation is needed prior to clinical implementation.