PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Tugba Akinci D'Antonoli, Lucas K. Berger, Ashraya K. Indrakanti, Nathan Vishwanathan, Jakob Weiß, Matthias Jung, Zeynep Berkarda, Alexander Rau, Marco Reisert, Thomas Küstner, Alexandra Walter, Elmar M. Merkle, Daniel Boll, Hanns-Christian Breit, Andrew Phillip Nicoli, Martin Segeroth, Joshy Cyriac, Shan Yang, Jakob Wasserthal

arXiv:2405.19492v2eess.IVcs.CV

TL;DR

Manual MRI segmentation is labor-intensive and variable, while sequence and protocol differences challenge automated generalization. This study trains and evaluates a model for 80 structures using MRI and CT data. On the MRI test set, the combined-data model outperformed the MRI-only model on Dice score and NSD, and was applied to an internal aging study.

  • Problem

    Manual MRI segmentation is time-consuming, labor-intensive, subject to interrater variability, and prone to errors, while sequence and protocol variation challenges automated generalization.

    Source: Introduction · p. 3

  • Method

    The study trained and evaluated an open-source segmentation model on MRI and CT data, using radiologist-reviewed annotations for 80 anatomic structures.

    Source: Introduction · p. 4, Data Annotation · p. 6

  • Results

    On the MRI test set, the combined MRI-and-CT model scored 0.862 vs. 0.845 on Dice score and 0.941 vs. 0.927 on NSD compared with the MRI-only model (p<.001).

    Source: Ablation study · p. 12

  • Takeaways & Limitations

    In an internal dataset of over 8,000 abdominal MRI patients, the model was used to identify correlations between age and organ volume.

    Source: Discussion · p. 16

    The retrospective design limits the study, and prospective evaluation is needed to strengthen its findings.

    Source: Discussion · p. 16

Abstract

from arXiv · show

Since the introduction of TotalSegmentator CT, there is demand for a similar robust automated MRI segmentation tool that can be applied across all MRI sequences and anatomic structures. In this retrospective study, a nnU-Net model (TotalSegmentator) was trained on MRI and CT examinations to segment 80 anatomic structures relevant for use cases such as organ volumetry, disease characterization, surgical planning and opportunistic screening. Examinations were randomly sampled from routine clinical studies to represent real-world examples. Dice scores were calculated between the predicted segmentations and expert radiologist reference standard segmentations to evaluate model performance on an internal test set, two external test sets and against two publicly available models, and TotalSegmentator CT. The model was applied to an internal dataset containing abdominal MRIs to investigate age-dependent volume changes. A total of 1143 examinations (616 MRIs, 527 CTs) (median age 61 years, IQR 50-72) were split into training (n=1088, CT and MRI) and an internal test set (n=55; only MRI), two external test sets (AMOS, n=20; CHAOS, n=20; only MRI), and an internal aging-study dataset of 8672 abdominal MRIs (median age 59 years, IQR 45-70) were included. The model showed a Dice Score of 0.839 on the internal test set and outperformed two other models (Dice Score, 0.862 versus 0.759; and 0.838 versus 0.560; p<.001 for both). The proposed open-source, easy-to-use model allows for automatic, robust segmentation of 80 structures, extending the capabilities of TotalSegmentator to MRIs of any sequence. The ready-to-use online tool is available at https://totalsegmentator.com, the model at https://github.com/wasserth/TotalSegmentator, and the dataset at https://zenodo.org/records/14710732.

Materials and Methods

The study develops an open-source model for sequence-independent MRI segmentation, addressing the variability and manual effort that complicate MRI segmentation. Its methods combine diverse MRI and CT training data, radiologist-reviewed annotations, model comparisons, and an internal aging-study application.

Training and internal test dataset

Data Annotation

  • All 616 MRI examinations received annotations reviewed and corrected by radiologists, with iterative nnU-Net training used to accelerate annotation.Two board-certified radiologists initiated and supervised annotation; final annotations served as the training and testing reference standard.Source: Data Annotation · p. 6

A model

  • The nnU-Net model was trained at 1.5-mm isotropic resolution, with a 3-mm version intended to reduce memory requirements.The 80 structures were split across six models; the 3-mm comparison used 50 selected structures.Source: A model · p. 6, Data Annotation · p. 6

Baseline Models

Ablation study

  • The annotated dataset, model code, and ready-to-use online tool are publicly available.The study also trained MRI-only and CT-only models for an ablation comparison with the combined MRI-and-CT model.Source: Ablation study · p. 7

Training dataset

  • Segmentation performance was assessed using Dice score and normalized surface distance, with Wilcoxon signed-rank tests for model comparisons.NSD measures how often surface distance is less than 3 mm; values for both metrics range from 0 to 1.Source: Training dataset · p. 7

Dataset Characteristics

  • 1,143 examinations were included, with 1,088 MRI and CT examinations for training and 55 MRI examinations in the internal test set.The MRI data varied in contrast, sequence, resolution, scanner, and acquisition site.Source: Dataset Characteristics · p. 8

Aging study dataset

B. Internal aging study dataset

TotalSegmentator MRI achieved strong segmentation scores across 80 structures and outperformed two public models on their supported subsets. The paper also reports age-related organ-volume correlations and discusses performance boundaries for MRI.

Segmentation Evaluation

  • Dice score was 0.862 vs 0.759 for MRSegmentator and 0.838 vs 0.560 for AMOS (both p<.001) on supported subsets; the all-80-structure Dice score was 0.839.The model’s all-80-structure NSD was 0.907.Source: Segmentation Evaluation · pp. 9–10
  • On external tests, TotalSegmentator MRI outperformed MRSegmentator on AMOS (Dice 0.886 vs 0.860) and CHAOS (0.890 vs 0.819); AMOS scored best on AMOS but worst on CHAOS.The proposed model and MRSegmentator comparisons were significant on both datasets; AMOS’s AMOS-set advantage did not generalize to CHAOS.Source: Segmentation Evaluation · p. 10
  • On the CT test set, TotalSegmentator CT led with Dice 0.970, followed by TotalSegmentator MRI at 0.966.TotalSegmentator MRI’s NSD was 0.996 versus 0.997 for TotalSegmentator CT.Source: Segmentation Evaluation · p. 10

Ablation study

  • On the MRI test set, the MRI+CT-trained model exceeded the MRI-only model (Dice 0.862 vs 0.845; NSD 0.941 vs 0.927), while CT-only training performed better on CT.For CT, the reported NSD was 0.971 for CT-only versus 0.966 for MRI+CT; both comparisons had p<.001.Source: Ablation study · p. 12

Failure cases

  • MRI image quality constrained segmentation: anisotropy, section thickness over 6 mm, and low contrast complicated small-structure identification, with failures shown for bowel, pancreas, iliac vessels, and metatarsals.Runtime and memory requirements are referred to the prior TotalSegmentator CT paper because the models share nnU-Net architecture and baseline resolutions.Source: Failure cases · p. 13, Figure 6 · p. 14, Runtime · p. 14

Evaluation of Age-related Differences

  • Age correlated positively with left and right adrenal volumes (rs=0.323 and 0.310) and negatively with kidney, liver, and spleen volumes (rs from −0.067 to −0.152).Figure 4 presents training-dataset parameter histograms described as demonstrating high diversity (561 MRIs and 527 CTs).Source: Figure 4 · p. 9, Evaluation of Age-related Differences · p. 14

Discussion

  • The paper extends TotalSegmentator to MRI of any sequence, makes the model and training data publicly available, and identifies organ-volume analysis as a potential application.The authors also describe potential use in existing clinical workflows and research projects.Source: Discussion · pp. 15–16
  • The retrospective study design limits the evidence; external datasets were included, but the authors say prospective evaluation is still needed.The authors also report lower performance on MRI than CT, attributing the difference to MRI heterogeneity and lower spatial resolution and signal-to-noise ratio.Source: Discussion · p. 16

S1. List of sources of external IDC datasets

The section lists external imaging-data sources, spanning cancer-imaging collections, clinical trial data, and other imaging datasets.

S1. List of sources of external IDC datasets

S2. List of all segmented structures

The supplement enumerates the segmented anatomy and identifies pretrained models used to create initial annotations for selected structures.

S2. List of all segmented structures

  • The segmented set spans abdominal and thoracic organs, vessels, bones, muscles, brain, and adipose tissue.The list names 80 structures, including paired left- and right-sided anatomy.Source: S2. List of all segmented structures · p. 20

S3. List of all pretrained models which were used during the data annotation

S4. Training details

Training used specified nnU-Net settings and divided the full 80-structure model into smaller models to manage memory; the supplement also lists comparison-set structures.

S4. Training details

  • Training used 3d_fullres, MR normalization, 2,000 epochs, no mirroring, and no ensembling.Mirroring was removed to help distinguish left from right structures; ensembling was omitted to keep runtime low.Source: S4. Training details · p. 20
  • The 80 structures were split into six parts, each trained with a smaller nnU-Net, because one model required very high memory.Source: S4. Training details · p. 20
  • TotalSegmentator MRI-3 combined 50 main structures in one model without unreasonable memory requirements.Source: S4. Training details · p. 20

S5. List of segmented structures for comparison to TotalSegmentator MRI-3 and MRSegmentator and AMOS model

S6. Supplemental Table

The supplemental table describes volume cutoffs and age-quartile tests, while the figures show dataset age, structure-level segmentation results, and age-related organ-volume patterns.

S6. Supplemental Table

  • The supplemental table reports structure-specific volume cutoffs and age-quartile p-values, using Kruskal-Wallis tests and a Bonferroni-adjusted significance threshold below 0.0001.Segmentations below the lower cutoff were excluded.Source: S6. Supplemental Table · p. 21

S7. Supplemental Figures

  • Figure S1 shows the age distribution of the aging-study dataset.Source: Figure S1 · p. 22
  • Figure S2 plots median Dice scores and NSD by anatomical structure, alongside individual-subject results.Source: Figure S2 · p. 23
  • Figure S3 shows correlations and age-quartile boxplots for volume changes in six abdominal organs.Source: Figure S3 · p. 24

S8. Supplemental References

The supplemental references include imaging datasets, annotations, and research publications.

Loading 2405.19492v2…