Semantic segmentation of human oocyte images using deep neural networks

Area of Science:

• Reproductive Medicine
• Artificial Intelligence in Healthcare
• Biomedical Image Analysis

Background:

• Infertility affects a significant portion of the global population.
• In vitro fertilization (IVF) is a key assisted reproductive technology (ART).
• Accurate assessment of oocyte and embryo quality is crucial for IVF success, but current morphological assessments are subjective.

Purpose of the Study:

• To develop and compare deep neural networks for the semantic segmentation of human oocyte images.
• To create an objective, automated method for assessing oocyte quality.
• To establish a foundation for new versions of predefined neural networks in medical image analysis.

Main Methods:

• Utilized deep learning, specifically convolutional neural networks (CNNs), for semantic oocyte segmentation.
• Compared the performance of various deep neural network architectures.
• Analyzed the merits and limitations of selected deep neural network models in a case study.

Main Results:

• Analyzed 71 deep neural network models.
• The DeepLab-v3-ResNet-18 model variant achieved the highest performance.
• Achieved approximately 85% training accuracy and 79% validation accuracy, with weighted intersection over union (wIoU) of 0.897 and global accuracy (gAcc) of 0.93 on test patterns.

Conclusions:

• The proposed deep learning approach enables the creation of highly accurate deep neural models for semantic oocyte segmentation.
• These models can serve as reliable predefined networks for other medical imaging tasks.
• The study demonstrates the potential of AI in objective gamete quality assessment for improved ART outcomes.