Integrating self-configuring and foundational deep learning segmentation models for identifying the anal sphincter complex and perianal fistulas on pelvic MRI

Perianal fistulas remain one of the most common complications associated with Crohn's disease, with an urgent need for improved interventional guidance and surgical planning. Pelvic magnetic resonance imaging (MRI) is routinely used for noninvasive imaging assessment of perianal fistulizing Crohn's Disease (CD-PAF), but suffers from significant inter-reader variability in accurately determining fistula tracts vis a vis anorectal anatomy. Towards overcoming these issues, we present a novel approach which integrates self-configuring segmentation frameworks (nnU-net) with generalized foundation models (MedSAM) toward the task of automated segmentation of fistula tracts and as well as internal and external sphincter muscles on pelvic MRI. We utilized a cohort comprising both baseline and follow-up MRI scans from 92 CD-PAF patients, for which manual annotations of all three anorectal structures were available. In hold-out validation, the integrated MedSAM-nnUnet model yielded the best overall performance in segmenting out the internal (Dice of 0.96 ± 0.18) and external sphincter muscles (Dice of 0.59 ± 0.09), as well as perianal fistulae (Dice of 0.55 ± 0.09); which represented significant improvements over MedSAM and nnU-net models individually. Integrating foundation with self-configuring segmentation models offers a novel automated annotation approach for detailed visualization of anorectal anatomy to guide surgical interventions in CD-PAF patients.