An intra- and inter-class context and consistency network for supervised and semi-supervised blastocyst segmentation.

Area of Science:

• Reproductive Medicine and Embryology
• Medical Image Analysis
• Artificial Intelligence in Healthcare

Background:

• Blastocyst quality is critical for embryo implantation potential.
• Accurate identification of blastocyst morphology is essential for objective assessment.
• Current segmentation methods may lack precision in capturing complex tissue structures.

Purpose of the Study:

• To explore semi-supervised learning (SSL) for enhanced blastocyst segmentation.
• To investigate structural associations between all blastocyst tissues.
• To improve the performance of blastocyst tissue segmentation.

Main Methods:

• Developed I2C2Net, a framework utilizing supervised and semi-supervised learning for automatic blastocyst segmentation.
• Incorporated Intra-Class Context Module (IACCM) and Inter-Class Context Module (IRCCM) for feature aggregation and morphological understanding.
• Introduced a Consistency Module (CM) for supervised training and a semi-supervised approach addressing data scarcity.

Main Results:

• I2C2Net achieved state-of-the-art performance in Accuracy, Precision, Recall, Dice, and Jaccard index compared to supervised methods.
• The semi-supervised I2C2Net demonstrated superior results over other SSL approaches, with significant gains in Accuracy, Precision, Dice, and Jaccard index.
• Ablation studies confirmed the effectiveness of the proposed IACCM, IRCCM, and CM modules.

Conclusions:

• I2C2Net effectively segments human embryo blastocysts, enhancing precision through its unique contextual modules.
• The semi-supervised version of I2C2Net offers a robust solution for clinical applications with limited annotated data.
• The framework demonstrates significant improvements in segmentation metrics, highlighting its potential for objective embryo quality assessment.