MRAnnotator: multi-anatomy and many-sequence MRI segmentation of 44 structures.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Anatomy

Background:

• Accurate segmentation of diverse anatomic structures in MRI is crucial for clinical diagnosis and research.
• Existing deep learning models may struggle with generalizability across different datasets and acquisition parameters.
• Developing robust multi-anatomy segmentation models is an ongoing challenge in medical image analysis.

Purpose of the Study:

• To develop and evaluate a deep learning model, MRAnnotator, for accurate multi-anatomy segmentation on diverse MRI scans.
• To assess the generalizability of the developed model across different clinical sites and imaging centers.
• To benchmark MRAnnotator against existing state-of-the-art MRI segmentation models.

Main Methods:

• A retrospective study utilizing two curated datasets: an internal dataset (1518 MRI sequences) and an external dataset (397 MRI sequences).
• Annotation of 44 anatomic structures using a model-assisted workflow with manual finalization.
• Training an nnU-Net model (MRAnnotator) on the internal dataset and evaluating its performance and generalizability on the external dataset.
• Benchmarking MRAnnotator against an AMOS-trained nnU-Net, TotalSegmentator MRI (TSM), and MRSegmentator (MRS) using Dice scores.

Main Results:

• MRAnnotator achieved an average Dice score of 0.878 on the internal dataset and 0.875 on the external dataset, indicating strong generalization.
• The model demonstrated comparable performance to an AMOS-trained nnU-Net on the AMOS test set (Dice 0.889 vs 0.895).
• MRAnnotator significantly outperformed TSM (Dice 0.822) and MRS (Dice 0.867), with P < .001 for both comparisons.

Conclusions:

• MRAnnotator provides robust and generalizable multi-anatomy segmentation for 44 structures in MRI.
• The model's performance surpasses existing methods like TSM and MRS.
• Future work will expand the model to include additional anatomic structures, with model weights publicly available.