Abstract
In In-Vitro Fertilization (IVF), the selection of high-quality embryos is a critical step for successful implantation. This process relies heavily on the morphological assessment of 5-day-old embryos, known as blastocysts. However, manual assessment is subjective and labor-intensive. This paper presents I2CNet, a novel deep-learning framework designed for high-precision automated segmentation of human blastocysts. By explicitly modeling intra-class consistency and inter-class anatomical relationships, I2CNet addresses the inherent challenges of low-contrast medical imaging, achieving state-of-the-art performance on benchmark datasets.
Introduction
The success of IVF treatments is largely dependent on the accurate selection of the most viable embryo. Embryologists typically grade blastocysts based on the development of specific tissues: the Inner Cell Mass (ICM), the Trophectoderm (TE), and the Blastocoel. Despite its importance, manual grading suffers from significant inter-observer variability and high workload. Automated segmentation offers a path toward objective, reproducible, and efficient embryo assessment. However, blastocysts present unique computer vision challenges due to their transparency and the low-light conditions required to maintain embryo viability during imaging.
Challenges in Blastocyst Segmentation
Automating blastocyst segmentation is hindered by two primary technical hurdles:
- Intra-class blur: Tissues such as the ICM often exhibit inconsistent textures across their surface, leading to "holes" or fragmented regions in standard segmentation outputs.
- Indistinct boundaries: The structural edges between adjacent tissues (e.g., the TE and the Zona Pellucida) are frequently faint and complex, causing traditional models like U-Net to fail at these critical borders.
The I2CNet Architecture
To overcome these issues, the authors propose I2CNet, an encoder-decoder architecture integrated with specialized I2C modules at multiple scales. The core innovation lies in two complementary components:
IntrA-Class Context Module (IACCM)
The IACCM addresses intra-class inconsistency by decomposing the multi-category segmentation task into individual binary tasks. By supervising the aggregation of features belonging to the same class, the module ensures that a tissue region is represented consistently across all its constituent pixels, effectively filling "holes" caused by texture variations.
InteR-Class Context Module (IRCCM)
The IRCCM leverages the biological "inside-out" hierarchy of the blastocyst (ICM → Blastocoel → TE → ZP). It captures the spatial correlations between adjacent tissues. A key feature of the IRCCM is a Weighted Mapping Function $U(\cdot)$, which redistributes probability weights to "stimulate" hard-to-classify pixels near boundaries. By focusing the network's attention on regions where confidence is low (near 0.5), I2CNet achieves significantly sharper and more accurate tissue borders.
The network further employs Deep Supervision, calculating loss and providing feedback at every scale of the decoder to ensure robust feature learning throughout the architecture.
Experimental Evaluation
The model was evaluated using the benchmark dataset introduced by Saeedi et al. (2017), which contains 249 human blastocyst images with expert-annotated ground-truth masks. The dataset was split into 199 images for training (80%) and 50 for testing (20%).
Results
I2CNet demonstrated superior performance compared to traditional CNNs (U-Net, DeepLabv3+) and modern Transformer-based models (Swin-Unet). Key metrics include:
- Accuracy: 94.14%
- Jaccard Index (IoU): 85.25%
- Dice Coefficient: 91.56%
Ablation studies confirmed the specific roles of the sub-modules: while the IACCM improved the internal consistency of tissue regions, the IRCCM was the primary driver for sharpening the difficult boundaries between layers.
Conclusion
I2CNet represents a significant advancement in automated embryo assessment. By incorporating biological priors and focusing on boundary uncertainty, it achieves high-precision segmentation that outperforms existing general-purpose architectures. While the current study is limited by the scale of available public datasets, it provides a robust foundation for future clinical tools that can assist embryologists in objective embryo grading and selection.
Summary updated