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

Area of Science:

• Medical Imaging
• Deep Learning
• Anatomical Segmentation

Background:

• Automated segmentation of 3D pelvic topography using deep learning networks can enhance preoperative planning and surgical navigation.
• Anisotropic MRI scans, common in clinical practice, have limitations due to uneven resolution, hindering accurate segmentation of pelvic structures.

Purpose of the Study:

• To compare the segmentation performance of deep learning models trained on anisotropic versus isotropic MRI reconstructions for pelvic muscle and nerve tissue.
• To evaluate the impact of MRI data isotropy on the accuracy and anatomical fidelity of 3D pelvic models.

Main Methods:

• Two nnU-Net models were trained using T2-weighted MRI scans from 35 rectal cancer patients: one on anisotropic data and another on isotropically reconstructed volumes.
• Model performance was assessed using Dice Similarity Coefficient (DSC), Intersection over Union (IoU), precision, and recall for segmenting sacral nerves, obturator nerve, and piriformis muscle.

Main Results:

• The model trained on isotropic MRI reconstructions demonstrated superior segmentation performance, particularly 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 analysis revealed smoother, more natural anatomical representations with isotropic models compared to the rougher surfaces from anisotropic models.

Conclusions:

• Isotropic MRI reconstructions provide improved segmentation accuracy and more anatomically correct 3D pelvic models.
• Favoring isotropic reconstruction in clinical workflows can enhance segmentation accuracy and efficiency for pelvic MRI, benefiting preoperative planning and surgical navigation.