Benchmarking of deep learning methods for generic MRI multi-organ abdominal segmentation.

Area of Science:

• Medical Imaging
• Deep Learning
• Computational Anatomy

Background:

• Deep learning has advanced abdominal CT segmentation but MRI segmentation remains challenging due to signal variability and annotation costs.
• Existing MRI segmentation models often use limited MRI sequences, potentially hindering their generalizability.
• Automated segmentation tools are crucial for efficient analysis of medical imaging data.

Purpose of the Study:

• To benchmark state-of-the-art, open-source abdominal MRI segmentation models.
• To evaluate a novel synthetic data-trained model (ABDSynth) for MRI segmentation.
• To assess model accuracy and generalizability across diverse datasets and imaging parameters.

Main Methods:

• Comprehensive benchmarking of three established open-source models: MRSegmentator, MRISegmentator-Abdomen, and TotalSegmentator MRI.
• Introduction and evaluation of ABDSynth, a SynthSeg-based model trained solely on CT segmentations.
• Performance assessment using three independent public datasets covering multiple manufacturers, MRI sequences, and acquisition variations.

Main Results:

• MRSegmentator demonstrated superior performance and generalizability among the evaluated models.
• ABDSynth achieved slightly lower accuracy but presents a viable option when manual annotation resources are constrained.
• Models trained on real, heterogeneous, multimodal data generally yielded the best outcomes.

Conclusions:

• Benchmarking reveals significant performance differences among open-source abdominal MRI segmentation tools.
• MRSegmentator is recommended for high-accuracy and generalizable abdominal MRI segmentation.
• ABDSynth offers a cost-effective alternative for segmentation tasks with limited training data availability.