A deep learning algorithm for automatic 3D segmentation and classification of the sheep placenta in magnetic...

Area of Science:

• Biomedical Imaging
• Veterinary Science
• Computational Biology

Background:

• Magnetic resonance imaging (MRI) offers advanced non-invasive assessment of placental morphology and physiology.
• Current placental MRI analysis in preclinical sheep models relies on manual segmentation, leading to time-consuming workflows and inter-operator variability.
• Standardized and automated analysis is crucial for reproducible preclinical research on placental development and function.

Purpose of the Study:

• To develop and validate the first fully automated MRI analysis pipeline for the ovine placenta.
• To leverage deep learning models for accelerated and standardized segmentation and classification of placental structures.
• To reduce operator bias and manual burden in preclinical placental assessment.

Main Methods:

• Evaluation of 2D and 3D self-configuring nnU-Net models against a conventional 2D U-Net for placental segmentation.
• Training and testing on expert-annotated ovine placental MRI datasets.
• Implementation of a YOLOv11-based detection module for placentome localization and morphological classification (Types A-D).

Main Results:

• The 3D nnU-Net achieved a Dice score >0.81, outperforming 2D methods and reducing processing time by over 95%.
• The automated pipeline closely matched manual segmentation contours.
• The YOLOv11 module achieved high performance metrics: 88.7% precision, 92.1% recall, mAP50 of 95.1%, and mAP50-95 of 93.5% for placentome classification.

Conclusions:

• A dual-stage, fully automated workflow for in vivo quantification of ovine placentome morphology has been established.
• This pipeline enables high-throughput, reproducible assessment of placental adaptation in preclinical models.
• The automated system reduces operator bias and manual effort, supporting MRI biomarker validation for placental health.