Abstract
Sacral neuromodulation (SNM) demands precise device placement, but current implantation methods do not utilize quantitative imaging measurements. Surgical fluoroscopic images are made routinely but used for only surgeon visual assessment. The images are challenging for automation of identification and measurement of surgical features due to the low-resolution images and extremely limited annotated datasets. Conventional semantic segmentation methods often fail to meet stringent medical accuracy due to (1) reliance on single-pass image evaluation and (2) noisy outputs requiring rigid post-processing that risks eroding true structures. Here we have developed an AI framework to address these issues. A CNN classification network was used to identify subregions containing medical devices, enhancing robustness by leveraging variations in contrast, brightness, and anatomy in subregions. An adaptive AI-driven noise reduction method eliminated the need for rigid post-processing, improving average IoU (Intersection Over Union) by 252.6% and achieving sub-millimeter precision in device localization relative to expert annotations.Clinical Relevance- Establishes AI-based methods for reducing noise and identification of relevant structures in surgical fluoroscopic images that will be useful for SNM implantation and other precise surgical needs as surgical technologies become more automated and quantitative.