TotalSegmentator MRI: Robust Sequence-independent Segmentation of Multiple Anatomic Structures in MRI

Area of Science:

• Medical imaging
• Artificial intelligence in radiology
• Anatomy segmentation

Background:

• Demand for automated MRI segmentation tools similar to TotalSegmentator CT.
• Need for a robust model applicable across all MRI sequences and anatomic structures.

Purpose of the Study:

• To develop and evaluate an automated MRI segmentation model for robust segmentation of major anatomic structures, independent of MRI sequence.
• To extend TotalSegmentator capabilities to MRI scans.

Main Methods:

• An nnU-Net model (TotalSegmentator MRI) was trained on 1143 MRI and CT scans to segment 80 anatomic structures.
• Performance was evaluated using Dice scores on internal and external test sets, compared against existing models and TotalSegmentator CT.
• The model was applied to a large dataset to investigate age-dependent organ volume changes.

Main Results:

• The model achieved a Dice score of 0.839 on the internal MRI test set, outperforming two other models.
• Performance on the TotalSegmentator CT test set (Dice score 0.966) closely matched TotalSegmentator CT (0.970).
• A strong correlation was found between age and organ volume in the aging study.

Conclusions:

• The proposed open-source TotalSegmentator MRI model enables automatic, robust segmentation of 80 structures across all MRI sequences.
• The tool is easy to use and extends TotalSegmentator's utility to MRI.
• Available online resources facilitate broad adoption and further research.