Blasto-Net: An Explainable Multi-Task Learning for Blastocyst Segmentation, Grading, and Implantation Prediction

In vitro fertilization (IVF) success depends heavily on the accurate selection of viable embryos. Current clinical assessment relies on manual morphological grading of blastocysts, which is inherently subjective and prone to inter-observer variability. This study aims to develop an automated, interpretable, and objective deep learning framework to segment blastocyst compartments (Inner Cell Mass, Trophectoderm, and Zona Pellucida), grade their morphology, and predict implantation potential.

The authors propose Blasto-Net, a multi-task learning architecture that processes images in a single forward pass. The model uses an EfficientNet-B3 encoder to extract hierarchical features, which are then processed by a UNet-style decoder. Key innovations include:

• Attention Mechanisms: Integration of the Convolutional Block Attention Module (CBAM) and a novel Edge-Aware Attention Module (EAAM) to sharpen boundary detection.
• Topology-Specific Heads: Specialized segmentation branches for each compartment, incorporating domain-specific guidance like Laplacian-of-Gaussian (LoG) filters for the compact ICM and Sobel edge maps for the thin-ring structures of the TE and ZP.
• Multi-Task Loss: A composite loss function that weights tasks based on their difficulty and anatomical constraints.
• Interpretability: Use of Grad-CAM++ to visualize the anatomical regions influencing the model's predictions, providing clinical transparency.

Blasto-Net demonstrates robust performance across all tasks. It achieves high Dice scores for segmentation (94.93% for ICM, 91.60% for ZP, and 88.82% for TE). For clinical prediction, the model achieves an F1-score of 80.0% for implantation outcomes, with a high recall of 94.12%, suggesting it is effective at identifying viable embryos. The inclusion of edge-aware modules significantly improves the delineation of thin, irregular structures that are typically difficult to segment.