Target Review Group: Clinical Radiologists and AI/Machine Learning Researchers in Medical Imaging
Abstract:
This paper introduces and validates TotalSegmentator MRI, a deep learning model based on the nnU-Net framework, designed for the automatic, robust, and sequence-independent segmentation of 80 major anatomic structures in Magnetic Resonance Imaging (MRI). Motivated by the success of TotalSegmentator CT, this retrospective study trained the model on a highly diverse cohort of 1143 examinations (616 MRIs, 527 CTs) spanning a variety of MRI sequences, contrasts, and scanner types, with annotations manually refined by board-certified radiologists. The evaluation on an internal MRI test set yielded a median Dice Score of 0.839 and a Normalized Surface Distance (NSD) of 0.907 across all 80 structures. The model significantly outperformed two publicly available baseline models (MRSegmentator and AMOS) across multiple internal and external test datasets (p<.001). Ablation analysis confirmed that training on combined MRI and CT data enhanced the model’s performance on MRI segmentation, demonstrating effective cross-modality data augmentation. Furthermore, application to an internal aging-study dataset (n=8672 abdominal MRIs) successfully demonstrated correlations between age and organ volume changes (e.g., negative correlation with liver and kidney volumes, positive correlation with adrenal gland volumes), underscoring the model's high utility for opportunistic screening and large-scale volumetric analysis. TotalSegmentator MRI provides an open-source, robust solution for complex volumetric tasks across diverse MRI sequences.
TotalSegmentator MRI: Robust Sequence-independent Segmentation of Multiple Anatomic Structures in MRI
Model Objective and Framework: The primary goal was to develop an automated segmentation tool, TotalSegmentator MRI, capable of robustly segmenting 80 major anatomic structures independent of the MRI sequence used. The model utilizes the nnU-Net framework, which automatically configures hyperparameters based on dataset characteristics.
Dataset Composition (Materials and Methods): The model was trained on a comprehensive retrospective dataset totaling 1143 examinations (616 MRIs and 527 CTs).
Diversity: The MRI training set (n=576) was randomly sampled from routine clinical studies over 12 years (2011–2023) to ensure high variability in contrast, section thickness, field strength, pulse sequences (T1w, T2w, PD), and acquisition sites (4 different sites, 30 different scanners).
Annotation: 80 structures were annotated, refined, and served as the reference standard, with initial segmentations generated either manually or using existing models, then iteratively corrected by board-certified radiologists.
Core Performance (Internal Test Set): On the internal MRI test set (n=55), the model achieved:
Dice Score: 0.839 [95% CI: 0.825, 0.851] across all 80 structures.
NSD: 0.907 [95% CI: 0.895, 0.919].
Resolution Effect: The higher resolution 1.5 mm model significantly outperformed the 3 mm resolution model (Dice Score 0.862 vs. 0.779; p<.001 for 50 main structures).
vs. MRSegmentator (40 structures): Dice Score 0.862 vs. 0.759 (p<.001).
vs. AMOS (13 structures): Dice Score 0.838 vs. 0.560 (p<.001).
vs. TotalSegmentator CT (CT test set, n=89): TotalSegmentator MRI closely matched the performance of the CT-specific model (Dice Score 0.966 vs. 0.970; p<.001).
Ablation Study Key Finding: Incorporating CT images into the training enhanced the model's performance on the MRI test set (Dice Score 0.862 vs. 0.845 for MRI-only training; p<.001), indicating that cross-modality training improves robustness.
Clinical Application (Aging Study): The model was applied to 8672 T1-weighted abdominal MRIs to analyze age-related volume changes:
Observed Correlations: Significant positive correlation was found between age and adrenal gland volume (rs ≈ 0.3), while significant negative correlation was found between age and kidney (rs ≈ -0.15), liver (rs = -0.096), and spleen (rs = -0.067) volumes (all p<0.0001).
Limitations and Failure Cases: Model performance was lower on MRIs than on CTs due to inherent MRI heterogeneity (e.g., high anisotropy, low contrast outside the area of interest), which challenges the detection of small structures (e.g., iliac arteries) and leads to errors like oversegmentation (colon) or missing parts (pancreas, small bowel).
Availability: The model, training dataset, and annotations are openly available: