LG-nnU-net for multilabel anal sphincter segmentation on MRI: quantitative evaluation in patients with anal fistula.

PURPOSE: To develop and evaluate a novel deep learning-based segmentation framework (LG-nnU-net) for multilabel segmentation of anal sphincter substructures on MRI, aimed at providing robust quantitative anatomical information without implying operative validation for clinical classification improvement.

METHODS

This single-center retrospective study, approved by the institutional ethics committee, included 272 patients diagnosed with anal fistula who underwent coronal T2-weighted MRI between January 2017 and December 2024. The dataset was divided into training and testing subsets in an 8:2 ratio, comprising 218 and 54 patients, respectively. Manual annotations of the levator ani muscle (LAM), puborectalis muscle (PRM), and superficial and subcutaneous parts of the external anal sphincter (EAS) were performed by three experienced abdominal radiologists. An optimized nnU-net architecture (LG-nnU-net) was implemented with asymmetric encoder expansion, Group Normalization, multi-scale feature aggregation, and deep supervision. Segmentation performance was compared to ResU-net, DenseU-net, and U-net++ using Dice similarity coefficient (DSC), Hausdorff distance (HD), and average symmetric surface distance (ASSD) for individual muscles, combined anatomical structures for quantitative evaluation only, and the entire sphincter complex.

RESULTS

LG-nnU-net achieved superior segmentation accuracy for all individual muscle structures: LAM DSC 76.89 % (95 % CI: 75.1-78.7 %), HD 9.45 mm, ASSD 0.74 mm; PRM DSC 76.71 % (95 % CI: 74.8-78.6 %), HD 11.11 mm, ASSD 0.71 mm; superficial EAS DSC 68.82 % (95 % CI: 66.3-71.4 %), HD 9.17 mm, ASSD 0.92 mm; subcutaneous EAS DSC 79.89 % (95 % CI: 77.9-81.9 %), HD 8.06 mm, ASSD 0.54 mm. For combined high-level structures (LAM + PRM) and low-level structures (superficial + subcutaneous EAS), DSCs were 79.94 % (95 % CI: 78.0-81.9 %) and 80.32 % (95 % CI: 78.5-82.1 %), respectively. The overall multi-structure segmentation of the anal sphincter complex achieved a DSC of 81.34 % (95 % CI: 79.6-83.1 %), HD 10.48 mm, and ASSD 0.46 mm, surpassing all comparison models.

CONCLUSION

The LG-nnU-net framework enables accurate, robust segmentation of anal sphincter substructures on MRI. These results indicate potential utility for quantitative analysis and preoperative assessment, but the clinical impact on high/low fistula classification remains unverified, requiring further multi-center validation and prospective outcome studies.