MRISegmenter: A Fully Accurate and Robust Automated Multiorgan and Structure Segmentation Tool for T1-weighted Abdominal MRI

Abstract

Background

There is a pressing demand to develop an automated segmentation tool for abdominal MRI that can provide accurate and robust segmentation in more than 60 abdominal organs and structures.

Purpose

To develop and evaluate the accuracy and robustness of an automated multiorgan and structure segmentation tool for T1-weighted abdominal MRI.

Materials and Methods

In this retrospective study, a T1-weighted abdominal MRI dataset composed of axial precontrast T1-weighted and contrast-enhanced T1-weighted arterial, portal venous, and delayed phases for each patient in a randomly selected sample was included at the National Institutes of Health Clinical Center. Each MRI series contained voxel-level annotations of 62 abdominal organs and structures. A three-dimensional segmentation (nnU-Net) model, called MRISegmenter, was trained on this dataset. This internal dataset was then randomly split into training and internal test sets. Evaluation was conducted on the internal test set and two external test sets (Abdominal Multi-Organ Segmentation Challenge 2022 [AMOS22] and Duke Liver). The predicted segmentations were compared against the radiologist-verified reference standard annotations using means ± SDs for the Dice similarity coefficient (Dice score) and normalized surface distance (NSD). The segmentation tool and dataset are publicly available at https://github.com/rsummers11/MRISegmenter.

Results

A total of 195 patients (training set, 135 patients [mean age, 54.7 years ± 16.3 {SD}; 72 male patients, 63 female patients]; internal test set, 60 patients [mean age, 51.1 years ± 14.4; 26 male patients, 34 female patients]) with 780 MRI scans containing 62 annotations each were included. On the internal test set, MRISegmenter achieved a mean Dice score of 0.861 ± 0.118 and a mean NSD of 0.924 ± 0.073. On external test sets AMOS22 (60 MRI scans) and Duke Liver (95 patients; 172 MRI scans), MRISegmenter attained a mean Dice score of 0.829 ± 0.133 and a mean NSD of 0.908 ± 0.067 and a mean Dice score of 0.933 ± 0.015 and a mean NSD of 0.929 ± 0.021, respectively.

Conclusion

MRISegmenter provided accurate and robust segmentation of 62 organs and structures at T1-weighted abdominal MRI.

© RSNA, 2025

Supplemental material is available for this article.

See also the editorial by Murphy in this issue.

Summary

MRISegmenter provided accurate and robust segmentation of 62 organs and structures, including 15 main organs, seven vessels, eight muscles, and 32 bones at T1-weighted abdominal MRI.

Key Results

• ■ MRISegmenter was trained and internally tested on a curated dataset of 195 patients and 780 T1-weighted abdominal MRI scans with voxel-level annotations of 62 abdominal organs and structures.

• ■ On the internal test set, MRISegmenter achieved a mean Dice score of 0.861 ± 0.118 (SD) and a mean normalized surface distance (NSD) of 0.924 ± 0.073; on the external test sets (ie, Abdominal Multi-Organ Segmentation Challenge 2022 and Duke Liver), MRISegmenter attained mean Dice scores of 0.829 ± 0.133 and 0.933 ± 0.015 and mean NSDs of 0.908 ± 0.067 and 0.929 ± 0.021, respectively.

• ■ MRISegmenter is publicly available at https://github.com/rsummers11/MRISegmenter.

Introduction

MRI is a fundamental and versatile tool in medical imaging, and it is widely incorporated into clinical workflows (1) due to its comprehensive diagnostic capabilities. Automating the delineation of organs and structures on MRI scans can enable downstream applications, such as the early detection of cancer (2), tracking interval changes in the size of radiologic findings, the diagnosis of diffuse and focal liver disease (3), radiation therapy planning and guidance (4), and quantification of body composition for opportunistic screening of diseases (5). Automated segmentation of multiple structures can improve patient outcomes without increasing the burden on radiologists (6).

Numerous studies have explored MRI segmentation across various body parts, such as the brain, heart, abdomen, and pelvis (3,7–9). Despite recent advancements, these studies focus on a limited set of organs and structures. In particular, multiorgan and structure segmentation for abdominal MRI lag behind the CT counterpart. Presently, the Abdominal Multi-Organ Segmentation Challenge 2022 (AMOS22) dataset (10) and TotalSegmentator MRI dataset (11) are the largest multiorgan and structure segmentation MRI datasets that are publicly available, to the knowledge of the authors. However, although the AMOS22 dataset focuses on abdominal organs, its annotations were provided for only 13 key abdominal organs across 60 patients. Additionally, its comprehensive data acquisition and detailed patient demographics information (or meta-information) were not available. Although the TotalSegmentator MRI dataset provides annotations for more organs and structures than AMOS22, its data were randomly drawn from the Picture Archiving and Communication System without detailed sequence type and contrast information. Meanwhile, the Duke Liver MRI dataset, which was recently released and has detailed meta-information, restricted the annotations provided solely to the liver (3). Therefore, to the knowledge of the authors, an abdominal MRI dataset with comprehensive annotations for various anatomic organs and structures (more than 60 abdominal organs and structures) and detailed meta-information do not exist currently. There is a pressing demand to provide such a dataset, which will also allow the creation of an accurate and robust automated segmentation tool for abdominal MRI.

Thus, the aim of the study was to develop and evaluate the accuracy and robustness of an automated multiorgan and structure segmentation tool for T1-weighted abdominal MRI.

Materials and Methods

Study Design and Patient Samples

This retrospective study was Health Insurance Portability and Accountability Act compliant and approved by the institutional review board at the National Institutes of Health Clinical Center. The requirement for written informed consent from the patients was waived.

To create the MRISegmenter internal dataset, the Picture Archiving and Communication System at the National Institutes of Health Clinical Center was queried for patients who underwent both abdominal MRI and CT on the same day between January 2019 and October 2021. Patients were excluded due to severe pathologic conditions that were challenging to annotate. Each patient had all four axial precontrast T1-weighted and contrast-enhanced T1-weighted arterial, portal venous, and delayed phases, respectively. CT images acquired on the same day for these patients were also collected, and these images included both contrast-enhanced and precontrast images. The CT images served as a base to provide pseudoannotations (artificial intelligence–generated annotations) for the MRI sequences.

For external test sets, AMOS22 and Duke Liver datasets were used. The AMOS22 dataset (ie, MRI subset) provided segmentation annotations for 13 main abdominal organs, including the kidneys, liver, and stomach. However, because AMOS22 does not provide detailed information regarding the sequence type and patient demographics, it is challenging to filter the dataset based on inclusion and exclusion criteria. Thus, no MRI scans were excluded. The Duke Liver dataset, containing multiparametric MRI scans and detailed data acquisition information, was primarily created for the automated classification of MRI sequences and exclusively contained liver segmentations. For the Duke Liver dataset, MRI scans without liver segmentation annotations and non–T1-weighted MRI scans were excluded.

Imaging Protocols

For MRI acquisition, all four axial T1-weighted phases were performed using the spectral adiabatic inversion recovery fat-suppressed technique with the three-dimensional volumetric interpolated breath-hold examination. Detailed MRI information, including repetition time and echo time, is provided in Table 1. CT acquisition parameters are provided in Appendix S1 and Table S1.

Table 1:

Abdominal MRI and Patient Characteristics in the Training and Internal Test Sets

Reference Standard

The internal MRI dataset was annotated to obtain voxel-level segmentation of 62 abdominal organs and structures, thereby offering comprehensive coverage across various anatomic regions within the abdomen (Fig 1). These structures were organized into four different categories based on their anatomic locations and physiologic structure: group 1 included 15 major organs, group 2 consisted of seven blood vessels, group 3 included eight muscles, and group 4 contained 32 skeletal structures. The annotations were obtained using a cross-domain cross-modality segmentation method (12) and an iterative learning method (13). For a detailed enumeration of the organs and structures within each group, see Appendix S2.

Figure 1:

MRISegmenter, a multiorgan and structure segmentation tool, segmented 15 major organs, seven vessels, eight muscles, and 32 skeletal structures using abdominal T1-weighted MRI studies. Visual examples show the segmentations for these organs and structures (images were produced by 3D Slicer, a free open-source software application [23]). IVC = inferior vena cava.

An iterative learning method was used to obtain comprehensive voxel-wise annotations for each of 62 structures for all MRI volumes across all included patients (12,13) (Fig S1). A detailed description of the annotation procedure is discussed in Appendix S3.

Verifying Annotations in the Training and Internal Test Sets

The internal MRI dataset was randomly split into a training and internal test set. A verification process was used to ensure the annotations matched the underlying anatomy of the organs and structures. If inaccurate annotations were identified, they were manually corrected by a research fellow (Y.Z.) with the guidance of a radiology resident (B. Khoury) and a senior board-certified radiologist (R.M.S., with >30 years of experience). ITK-SNAP (version 4.0; University of Pennsylvania), a manual annotation tool, was used for annotation (14). Specifically, within the training set, the radiology resident verified a subset in the training set. After this, the research fellow manually corrected the annotations and reviewed and amended the annotations for the rest. For the internal test set, both the radiologist and the radiology resident manually reviewed two sequences per patient. The research fellow then manually refined the annotations for all internal test set volumes based on their collective feedback.

Deep Learning Model

Segmentation framework.—MRISegmenter used the self-configuring three-dimensional nnU-Net segmentation framework (15,16). A three-dimensional full-resolution nnU-Net was trained with fivefold cross-validation on the training set, and it resulted in five models that were used in an ensemble for prediction on test sets. The number of training epochs was set to 2000 to account for the large number of structures. The loss function was an equally weighted combination of binary cross-entropy and soft Dice losses. Optimization of this loss function was achieved using the Adam optimizer with an initial learning rate of 10 ^e−2, a polynomial learning rate scheduler, and a batch size of two. The MRI volumes were preprocessed by following the standard nnU-Net preprocessing steps. The experiments were performed using an 80-GB graphics processing unit (A100; Nvidia) on the National Institutes of Health Clinical Center high-performance computing cluster. The inference time for one MRI is about 50 seconds using an A100 graphics processing unit. The segmentation tool is publicly available at https://github.com/rsummers11/MRISegmenter.

Experimental Setup and Datasets.—MRISegmenter’s performance was evaluated on both the internal and external test datasets. For the internal test set, segmentation results for all 62 organs and structures were analyzed. An additional experiment was conducted to compare MRISegmenter against the nnU-Net model trained on the AMOS22 dataset as well as two recent studies: TotalSegmentator-MRI and TotalVibeSegmentator (11,17). Two external test sets were used for external testing. Details regarding the datasets and models used for external testing are included in Appendix S4.

Statistical Analysis

The segmentation accuracy of the deep learning model, MRISegmenter, was quantitatively measured using the Dice similarity coefficient, also known as the Dice score, and normalized surface distance (NSD). Dice scores measure the overlap between the standard reference annotation and the predicted segmentation. NSD focuses on the agreement between the standard reference surface and the predicted segmentation surface with a predefined tolerance. For each patient, the means ± SDs for the Dice score and the NSD were calculated for each of the 62 structures. Three statistical tests were conducted to measure the tool’s performance using software (R, version 4.4.1; R Foundation for Statistical Computing). First, two-sided paired t tests were used to compare the Dice score and the NSD between MRISegmenter and other segmentation approaches. Second, a Welch t test was used to compare the Dice score and the NSD between male and female patients. P < .05 was indicative of a statistically significant difference. Third, the performance across different MRI scans was compared to investigate the segmentation consistency across all sequences. To account for the correlations between two series from the same patient, a mixed-effects statistical model was run with the Dice scores as the dependent variable, the MRI sequence type and organ as the fixed effects, and a random effect for each patient. Marginal means for each type of sequence and standard errors were obtained.

Results

Dataset Characteristics

Of the 632 patients who met inclusion criteria, 225 were randomly selected to be included in the study; 30 of these patients were excluded due to severe pathologic conditions that were challenging to annotate. This exclusion resulted in a total of 195 unique patients with 780 MRI scans, including both precontrast T1-weighted and contrast-enhanced T1-weighted arterial, portal venous, and delayed phases, totaling 69 248 two-dimensional sections. This internal MRI dataset was then randomly split into training (135 patients; mean age, 54.67 years ± 16.27 [SD]; 72 male patients, 63 female patients; 540 MRI scans) and internal (60 patients; mean age, 51.12 years ± 14.39; 26 male patients, 34 female patients; 240 MRI scans) test sets. Figure 2 shows the inclusion and exclusion criteria for the MRISegmenter training and internal test sets. Table 1 summarizes the acquisition parameters and demographic details of the training and internal test sets, including self-reported race and ethnicity information. Table 2 lists the patients’ medical history used in this study.

Figure 2:

Standards for Reporting of Diagnostic Accuracy Studies chart shows the inclusion and exclusion criteria for the internal dataset and the Duke Liver dataset, an external test set. NIH = National Institutes of Health.

Table 2:

Medical History for Patients Included in the Abdominal MRI Dataset Used for Training and Internal Testing

External testing of MRISegmenter was conducted on the AMOS22 MRI data subset (60 patients, 60 MRI scans) and the Duke Liver dataset (95 patients, 172 MRI scans). Figure 2 also shows the inclusion and exclusion criteria for the Duke Liver test set. Table S2 shows the corresponding acquisition parameters and demographic details. Additionally, the performance of MRISegmenter was compared with other state-of-the-art methods on the official validation set of the AMOS22 MRI subset (20 patients, 20 MRI scans).

Internal Testing

Overall, MRISegmenter achieved a mean Dice score of 0.861 ± 0.118 and a mean NSD of 0.924 ± 0.073 for all structures in all series in the internal test set. The segmentation results were also analyzed for each group of structures. The mean Dice score and mean NSD for group 1 (15 major organs) were 0.918 ± 0.104 and 0.959 ± 0.040, group 2 (vessels) were 0.913 ± 0.063 and 0.972 ± 0.039, group 3 (muscles) were 0.929 ± 0.079 and 0.971 ± 0.038, and group 4 (bones) were 0.806 ± 0.113 and 0.886 ± 0.075, respectively. Figure 3 shows the mean Dice score for the 62 organs and structures across all patients. Tables S3 and S4 show the Dice scores and the NSDs for all organs and structures.

Figure 3:

Dice scores for all 62 organs and structures (ranked in descending order of performance). The number of volumes in the training data subset that the structure was present in is indicated in parentheses.

Figure 4 shows the segmentation results for a patient in the internal test set in the axial, sagittal, and coronal views. The color map of the segmentation mask for each organ is shown in Figure S5.

Figure 4:

Images show reference standard (radiologist-verified annotations) and predicted segmentation (predict) results for one patient in the internal test set. Results are shown in the axial (A), sagittal (B), and coronal views (C). (A) Axial images show segmentation results from superior to inferior (left to right). (B) Sagittal images show segmentation results from the patient’s right to left. (C) Coronal images show segmentation results from anterior to posterior (left to right). The color map of segmentation for each organ is provided in Figure S5.

The mean Dice scores for the precontrast, arterial, venous, and delayed phases were 0.863 ± 0.231, 0.874 ± 0.217, 0.880 ± 0.212, and 0.876 ± 0.220, respectively (n = 60). The mean NSD values were 0.922 ± 0.208, 0.932 ± 0.196, 0.937 ± 0.191, and 0.933 ± 0.197, respectively (n = 60). Precontrast T1-weighted MRI exhibited the lowest performance, primarily because it is a precontrast sequence in which some organs (eg, pancreas) and structures (eg, vessels) do not show any enhancement patterns, which affects the precise segmentation of these structures. The mixed-effects model analysis found that the differences between precontrast and the arterial, venous, and delayed phases were statistically significant (n = 60; precontrast vs arterial, P = .02; precontrast vs venous, P < .001; precontrast vs delayed, P = .002), indicating a higher segmentation accuracy in postcontrast phases. However, no evidence of significant difference was observed among the arterial, venous, and delayed sequences (arterial vs venous, P = .89; arterial vs delayed, P > .99; venous vs delayed, P > .99), which suggested a comparable segmentation performance for these three sequences. Furthermore, there was no evidence of a significant difference between male and female patients for mean Dice score (P = .10) and NSD (P = .20).

The proposed tool was also compared against TotalSegmentator-MRI, TotalVibeSegmentator, and the nnU-Net model trained on the AMOS22 dataset. Appendix S4 and Table S5 describe these three recent segmentation models and their comparisons with MRISegmenter. Overall, MRISegmenter significantly outperformed these three models based on paired t tests (P < .001) (n = 60). The mean Dice score and mean NSD value across shared organs for each segmentation model were as follows: MRISegmenter, 0.918 ± 0.104 and 0.959 ± 0.040; TotalSegmentator-MRI, 0.792 ± 0.143 and 0.821 ± 0.072; TotalVibeSegmentator, 0.738 ± 0.183 and 0.750 ± 0.114; and nnU-Net model trained on the AMOS22 dataset, 0.827 ± 0.156 and 0.893 ± 0.092, respectively. Specifically, MRISegmenter outperformed the other methods by a large margin for most organs, such as the gallbladder, stomach, and duodenum (Table S5). The performance differences were subtle for large organs, such as the kidneys and the liver (Table S5). One contributing factor to the decreased segmentation performance of other methods was the presence of various pathologic conditions, such as multifocal liver tumors and pancreatic cysts, in the internal dataset. Figure 5 shows several instances in which other methods did not segment regions affected by pathologic changes, leading to decreased performance in these cases.

Figure 5:

Images show comparison of MRISegmenter, nnU-Net-AMOS, TotalVibeSegmentator, and TotalSegmentator-MRI with the reference standard on segmenting structures with different types of pathologic abnormalities. Images show alternative methods that struggled to segment the same regions (arrows) in (A) a patient with heterogeneous splenic attenuation at arterial phase enhancement (B) a patient with a hypointense kidney lesion, and (C) a patient with metastatic disease with multiple liver lesions.

External Testing

AMOS22 external test set.—Figure 6 shows the segmentation performance of MRISegmenter for each of the 13 organs across all 60 patients in the AMOS22 dataset. Overall, MRISegmenter achieved a mean Dice score of 0.829 ± 0.133 across 13 organs. On the official validation set of the AMOS22 MRI subset (20 patients, 20 volumes), MRISegmenter attained a mean Dice score of 0.844, which was the second-best performance compared with other methods that were directly trained on the AMOS22 dataset (CoTr, 0.775; nnFormer, 0.806; Swin-UNetr, 0.757; Unet, 0.856; UNetr, 0.753; and VNet, 0.837) (10). This observation is remarkable because MRISegmenter was not directly trained with the AMOS22 dataset. Figure S2 shows the segmentation results in a patient from the AMOS22 dataset on the axial, sagittal, and coronal planes. Figure S3 shows the segmentation results of a patient at coronal T1-weighted MRI in AMOS22, the external test set.

Figure 6:

Plot shows the segmentation performance (mean Dice scores ± SDs) of MRISegmenter for 13 major organs in the Abdominal Multi-Organ Segmentation Challenge 2022 dataset, an external test set. IVC = inferior vena cava.

Duke Liver external test set.—In the Duke Liver external test set, the mean Dice scores for the precontrast, arterial, and venous phases were 0.912 ± 0.055, 0.946 ± 0.012, and 0.942 ± 0.019, respectively. The mean NSD scores were 0.903 ± 0.070, 0.929 ± 0.024, and 0.956 ± 0.017, respectively. MRISegmenter achieved Dice scores exceeding 0.9 across all three sequences (precontrast, arterial, and venous). The lowest Dice score was observed in the precontrast sequence, attributable to the reduced contrast inherent in precontrast imaging. Figure S4 shows the segmentation results for a patient from the Duke Liver dataset in the axial, sagittal, and coronal planes.

Discussion

There is a pressing demand to develop an automated abdominal MRI segmentation tool that can provide accurate and robust segmentation of more than 60 abdominal organs and structures. Our proposed model, MRISegmenter, a fully automated multiparametric segmentation tool, was trained and internally tested on a curated dataset of 195 patients and 780 T1-weighted abdominal MRI volumes with voxel-level annotations of 62 abdominal organs and structures. In the internal test set, MRISegmenter achieved a mean Dice score of 0.861 ± 0.118 and a mean normalized surface distance (NSD) of 0.924 ± 0.073, significantly outperforming TotalSegmentator-MRI and TotalVibeSegmentator (P < .001). In the two external test sets, Abdominal Multi-Organ Segmentation Challenge 2022 (AMOS22) and Duke Liver, MRISegmenter attained mean Dice scores of 0.829 ± 0.133 and 0.933 ± 0.015 and mean NSDs of 0.908 ± 0.067 and 0.929 ± 0.021, respectively. Empirical results demonstrated that the model achieved mean Dice scores exceeding 0.90 for 15 major organs (except the adrenal glands), seven blood vessels, and eight muscles and 0.80 for 35 bones. Furthermore, we used the two large external test sets composed of abdominal MRI data from different institutions to show its robust performance. MRISegmenter obtained a mean Dice score of 0.829 ± 0.133 for 13 organs on the AMOS22 external test set and attained Dice scores exceeding 0.90 across all three sequences on the Duke Liver external test set.

The segmentation performance on AMOS22 was not as high as the results observed in the internal test set. MRISegmenter achieved a Dice score exceeding 0.85 for the majority of organs, with the exception of several smaller organs such as the gallbladder, esophagus, adrenal glands, and duodenum. This discrepancy can be attributed to the non–T1-weighted and nonaxial volumes in the AMOS22 dataset. These series types were unseen during the training phase of MRISegmenter and diverged from the training data distribution. Nevertheless, MRISegmenter showed its robustness to different coronal volumes because the nnU-Net framework incorporates a resampling procedure in the preprocessing stage. On the official validation set of the AMOS22 MRI subset, though MRISegmenter was not directly trained with the AMOS22 dataset, MRISegmenter still came in second compared with other state-of-the-art methods that were trained directly on the AMOS22 training set. This result showcases the robust segmentation capabilities of MRISegmenter on a dataset obtained from an external institution.

Furthermore, the segmentation performance was suboptimal for ribs and vertebrae, particularly in the upper and lower abdomen near transitional zones of the chest and pelvis. Results from the coronal volumes in the AMOS22 external test set contained a portion of the lower chest and showed that parts of the lung, ribs, and vertebrae were erroneously segmented. These errors occurred because the chest MRI scans were not included in the training data. Chest MRI studies are infrequently acquired in our institution because patient motion from breathing can corrupt the MRI sequences.

To the best of our knowledge, multistructure segmentation on the scale of our work has been explored in only a few recent studies. Chen et al (18) proposed a two-dimensional U-Net–based method to segment 10 major abdominal organs and bones using their dataset of 102 patients, but this study covered only a limited number of structures. Ji et al (10) provided the first publicly available MRI dataset containing multistructure labels as part of the AMOS22 challenge, and it featured annotations for 13 structures across 60 MRI volumes. However, the limited number of patients in this dataset posed substantial challenges for model generalization and lacked data acquisition parameters and patient demographic information. The Duke Liver external test set (3), although it provided detailed sequence information, was primarily released for automated classification of MRI sequences and exclusively contains liver segmentations. More recently, several studies on multiorgan and structure segmentation for MRI have emerged. Gu et al (19) proposed a universal bone segmentation network to segment skeletal structures on MRI scans. In parallel, Zhou et al (20) proposed the MRAnnotator to segment 49 structures in the whole body. Geißler et al (21) developed a multiorgan segmentation tool for T1-weighted Dixon MRI using global intensity nonlinear data augmentation. Akinci D’Antonoli et al (11) proposed the TotalSegmentator MRI, a versatile tool that not only segments most major anatomic structures at MRI, but also supports a wide range of MRI sequences. Häntze et al (22) proposed the MRSegmentator, which was able to delineate 40 organs and structures at MRI and CT. Graf et al (17) developed a TotalSegmentator-like tool for volumetric interpolated breath-hold MRI scans. However, these tools either encompassed fewer structures compared with our work, the datasets or models were not publicly available, or the datasets lacked detailed data acquisition and patient medical history. Furthermore, we quantitatively and empirically demonstrated that MRISegmenter is superior to other recent studies. In addition, our proposed dataset contains both data acquisition (eg, sequence type and contrast information) and detailed patient information (eg, demographics and medical history). It is worth noting that the proposed segmentation model and dataset would impact a wide range of clinical applications that require segmentation in their preprocessing pipeline (eg, body composition estimation, volumetric and quantitative analysis in disease diagnosis, and radiation therapy planning and treatment). It will not only address the shortage of publicly available MRI datasets with dense annotations but also allow more researchers to leverage the proposed tool and explore various clinical and translational applications, thus advancing abdominal radiology using MRI.

Our study had limitations. First, the internal dataset (training set, internal test set) included patients with a broad range of pathologic abnormalities, and an analysis of this effect posed on the segmentation accuracy of MRISegmenter has not been evaluated. Second, this study solely focused on fat-suppressed precontrast and dynamic contrast-enhanced T1-weighted MRI scans. Future work will extend the dataset to include other MRI sequences, such as T1- and T2-weighted images. Third, though using artificial intelligence–generated pseudolabels in the iterative learning process facilitated the annotation process and reduced annotation efforts, it also introduced annotation bias. Fourth, the patients who underwent both CT and MRI on the same day were selected, and patients with severe pathologic conditions were excluded. These inclusion and exclusion criteria were considered from a practical perspective for the annotation. Using CT images facilitated the annotation process and reduced annotation efforts. Patients with severe pathologic conditions were excluded because annotation is challenging; for example, in patients with many tumors and lesions, the boundaries between organs and structures were diffused and blurred, and thus difficult to annotate. Future work will extend the dataset to include other MRI sequences, such as other T1-weighted and T2-weighted images, as well as to recruit more radiologists and perform interrater variability studies for the segmentation results.

In conclusion, MRISegmenter, a fully automated multiparametric segmentation tool, provides accurate and robust segmentation of 62 organs and structures, including 15 main organs, seven vessels, eight muscles, and 32 bones at T1-weighted abdominal MRI. MRISegmenter is an efficient and accurate segmentation tool that can benefit various downstream clinical tasks.