MiMics-Net: A Multimodal Interaction Network for Blastocyst Component Segmentation

Area of Science:

• Reproductive medicine
• Artificial intelligence in healthcare
• Medical image analysis

Background:

• Global infertility rates are rising, increasing demand for effective assisted reproductive technologies.
• In vitro fertilization (IVF) success relies heavily on accurate blastocyst assessment, a currently manual, subjective, and error-prone process.
• Existing AI segmentation methods struggle with challenging blastocyst image conditions like low contrast and texture similarity.

Purpose of the Study:

• To develop a novel AI-based segmentation network for accurate blastocyst component analysis.
• To address limitations of current methods in handling complex image data and improve IVF success prediction.

Main Methods:

• Developed MiMics-Net, a lightweight multimodal interaction segmentation network.
• Utilized a multimodal blastocyst stem processing photometric intensity, local textures, and directional orientation.
• Incorporated dual-path grouped blocks and a lightweight refinement decoder for enhanced feature processing and spatial restoration.

Main Results:

• MiMics-Net achieved a Jaccard index score of 87.9% on a human blastocyst dataset.
• The model requires only 0.65 million trainable parameters, indicating computational efficiency.
• The multimodal approach and network architecture improved segmentation performance over existing methods.

Conclusions:

• MiMics-Net offers a promising AI solution for accurate blastocyst segmentation in IVF.
• The developed network effectively handles challenging image conditions, paving the way for more reliable IVF outcomes.
• This lightweight and efficient model can enhance the precision of assisted reproductive technologies.