Deep Learning Segmentation of Pelvic Soft Tissue in Isotropic and Anisotropic MRI Using Routine T2 Scans

Automated segmentation of 3D pelvic topography using deep learning networks has the potential to improve the accuracy of preoperative planning and surgical navigation. However, anisotropic MRI scans-often used in clinical practice-present a major limitation due to their uneven resolution across anatomical planes. We compared the performance of a model trained on anisotropic versus isotropic MRI reconstructions for the segmentation of pelvic muscle and nerve tissue.

A total of 35 rectal cancer patients were included, each with axial, sagittal, and coronal T2-weighted MRI scans. Isotropic volumes were reconstructed from anisotropic scans. Two nnU-Net models were trained: one on anisotropic data and the other on isotropic reconstructions. Both models were evaluated for their segmentation of the sacral nerves, obturator nerve, and piriformis muscle. Dice Similarity Coefficient (DSC), Intersection over Union (IoU), precision, and recall were calculated to assess model performance.

The performance of the model trained on isotropic images had better segmentation performance, with particularly noticeable improvement for the obturator nerve (DSC: 0.76 vs. 0.71). Both models achieved high DSC scores for the piriformis muscle (>0.95) and good performance for the sacral nerves (>0.86). Qualitative evaluation showed smoother and more natural anatomical representations in isotropic models, while anisotropic models displayed rougher, discontinuous surfaces.

Isotropic MRI reconstructions yielded slightly better segmentation results and provided more anatomically accurate 3D models. This suggests that isotropic reconstruction should be favored in clinical workflows to improve segmentation accuracy and efficiency in pelvic MRI images for accurate preoperative planning and surgical navigation.