Researchers at the National Institutes of Health Clinical Center have developed MRISegmenter, an automated segmentation tool that demonstrates high accuracy and robustness in segmenting over 60 abdominal organs and structures on T1-weighted MRI.

"MRISegmenter, a fully automated multiparametric segmentation tool, provides accurate and robust segmentation of 62 organs and structures, including 15 main organs, 7 vessels, 8 muscles, and 32 bones at T1-weighted abdominal MRI," reported Yan Zhuang, PhD, of the Department of Diagnostic, Molecular, and Interventional Radiology, Windreich Department of Artificial Intelligence and Human Health, Icahn School of Medicine at Mount Sinai, New York, NY, and colleagues.

This innovation addresses a longstanding clinical and research need for reliable, large-scale organ segmentation in abdominal imaging, noted investigators, Zhuang and colleagues reported.

Accurate segmentation of abdominal organs and structures is essential for diagnostic radiology, treatment planning, and research. However, manual segmentation is labor-intensive, subjective, and time-consuming. With over 60 abdominal organs and structures requiring delineation, there has been a growing demand for an automated, scalable solution capable of handling the complexity and variability inherent to abdominal MRI.

Development of MRISegmenter

The team retrospectively assembled a comprehensive T1-weighted abdominal MRI dataset, including axial precontrast and multiphase contrast-enhanced sequences. Each image set was meticulously annotated at the voxel level for 62 distinct anatomical structures.

Leveraging the versatile nnU-Net framework, the researchers developed MRISegmenter, a three-dimensional deep learning model trained on this richly annotated dataset. The internal data, comprising 195 patients and 780 MRI scans, was split into training (135 patients) and internal testing (60 patients) subsets. Further evaluation utilized two external datasets: the Abdominal Multi-Organ Segmentation Challenge 2022 (AMOS22) and the Duke Liver dataset.

The performance of MRISegmenter was measured by comparing its segmentations against radiologist-verified reference standards, using the Dice similarity coefficient (Dice score) and normalized surface distance (NSD) as primary metrics.

Results

On the internal test set, MRISegmenter achieved an impressive mean Dice score of 0.861 ± 0.118 and a mean NSD of 0.924 ± 0.073, indicating a high degree of overlap and boundary accuracy compared to manual annotations.

Performance on external datasets was equally promising:

• AMOS22 (60 MRI scans): Mean Dice score of 0.829 ± 0.133, NSD of 0.908 ± 0.067.

• Duke Liver (172 MRI scans from 95 patients): Mean Dice score of 0.933 ± 0.015, NSD of 0.929 ± 0.021.

These results underscore MRISegmenter's generalizability and robustness across different patient populations and imaging protocols, noted investigators.

Implications and Future Directions

The release of MRISegmenter, alongside its comprehensive training dataset, at MRISegmenter GitHub Repository marks an important milestone for the radiology community, noted investigators. Automated, accurate segmentation can facilitate more efficient workflows, improve reproducibility in research studies, and lay the groundwork for advanced radiomics and machine learning applications.

Future research may focus on expanding MRISegmenter's capabilities to other MRI sequences and integrating it into clinical decision support systems to further streamline abdominal imaging interpretation.

Conflicts of interest can be found in the published study.

Source: Radiology