Federated Task-Adaptive Learning for Personalized Selection of Human IVF-derived Embryos

Guangyu Wang, Tianrun Gao, Yuning Yang, Kai Wang, Yuanxu Gao, Li-Shuang Ma, Lei Chen, Guangdong Liu, Ping Zhang, Xiaohong LiuResearch Square 2024 · DOI

Abstract

In-vitro fertilization (IVF) offers a solution for couples facing infertility challenges. However, the success of IVF, particularly in achieving live-birth outcomes, heavily depends on embryologists to conduct morphological assessments of fertilized embryos, a process that is both time-consuming and labor-intensive. While artificial intelligence (AI) has gained recognition for its potential to automate embryo selection, the application of deep learning (DL) is constrained by privacy concerns associated with the requirement for centralized training on extensive datasets. In this paper, we have developed a distributed DL system, termed ‘FedEmbryo’, tailored for personalized embryo selection while preserving data privacy. Within FedEmbryo, we introduce a Federated Task-Adaptive Learning (FTAL) approach with a hierarchical dynamic weighting adaption (HDWA) mechanism. This approach first uniquely integrates multi-task learning (MTL) with federated learning (FL) by proposing a unified multitask client architecture that consists of shared layers and task-specific layers to accommodate the single- and multi-task learning within each client. Furthermore, the HDWA mechanism mitigates the skewed model performance attributed to data heterogeneity from FTAL. It considers the learning feedback (loss ratios) from the tasks and clients, facilitating a dynamic balance to task attention and client aggregation. Finally, we refine FedEmbryo to address critical clinical scenarios in the IVF processes, including morphology evaluation and live-birth outcomes. Experimental results indicate that FedEmbryo outperforms both locally trained models and state-of-the-art (SOTA) FL methods. Our research marks a significant advancement in the development of AI in IVF treatments.

Introduction

In-vitro fertilization (IVF) is a beacon of hope for millions, yet its success is heavily reliant on the "expert eye" of embryologists who manually grade embryos based on their shape and development. This process is subjective, labor-intensive, and prone to human error. While Deep Learning (DL) could automate this, medical data is notoriously difficult to centralize due to strict privacy regulations and the sensitive nature of human reproductive data.

Standard Federated Learning (FL)—where a global model is trained across multiple decentralized servers without ever seeing the raw data—offers a solution. However, FL often stumbles in the real world because different clinics have "heterogeneous" data (different patient demographics, imaging equipment, or clinical protocols). Furthermore, IVF involves assessing multiple morphological metrics alongside predicting the ultimate goal: a live birth.

The FedEmbryo Framework

To address these challenges, we developed FedEmbryo, a distributed system built on a Federated Task-Adaptive Learning (FTAL) approach. This framework enables decentralized collaboration of embryo data to build personalized selection models while preserving data privacy.

Overview of the FedEmbryo framework and the Federated Task-Adaptive Learning (FTAL) approach.

The core innovations include:

  • Unified Multitask Architecture: Each clinic uses a model with shared convolutional layers (ResNet-50) for general feature extraction and task-specific heads for individual metrics, supporting both single- and multi-task learning.
  • Hierarchical Dynamic Weighting Adaption (HDWA): This mechanism uses learning feedback (loss ratios) to dynamically adjust weights. At the intra-client level, it balances attention between different tasks (e.g., focusing more on "harder" tasks like fragmentation prediction). At the inter-client level, it ensures the global model learns effectively from both large and small data centers without bias.
  • Multimodal Fusion: For live-birth prediction, the model integrates embryo images with clinical metadata (e.g., maternal age, BMI, hormone levels) using a Multilayer Perceptron (MLP) branch to provide a more holistic prediction.

Datasets and Validation

The study utilized a massive multimodal dataset from the China Consortium of Assisted Reproductive Technology Investigation (CC-ARTI):

  • Training/Internal Testing: Four distinct clinical "clients" (A, B, C, and D) with patient counts ranging from 431 to 2,203 and over 20,000 embryo images.
  • External Validation: Two independent cohorts, Cohort E (992 patients) and Cohort F (479 patients), were used to test how well the model generalizes to entirely unseen environments.

Key Results and Insights

FedEmbryo consistently outperformed local training and state-of-the-art FL methods such as FedAvg and FedProx. In morphology assessment, the model saw average improvements of 20.31% in AUC for abnormal pronuclear detection and 35.48% for blastocyst formation prediction. For live-birth prediction, it achieved an AUC of 0.85 on internal sets and 0.76 on external sets, representing an 11% and 5% improvement over locally trained models, respectively.

Performance comparison of FedEmbryo against baseline methods across various clinical tasks.

Key insights from the study include:

  • The Power of Feedback: The HDWA mechanism proves that "listening" to the model's struggle is better than fixed weighting. By identifying which tasks or clinics are lagging based on loss ratios, the system self-corrects.
  • Beyond the Image: Significant performance jumps occurred when clinical factors like maternal age and endometrial thickness were added, highlighting the importance of multimodal data.
  • Visualizing Logic: Saliency maps (Integrated Gradients) showed that FedEmbryo focuses on clinically relevant features, such as nuclei on Day 1 and cell quantity on Day 3, validating its decision-making process.

Conclusion

While FedEmbryo marks a significant advancement, limitations remain, including the need for validation across more diverse racial and ethnic groups and the computational overhead of the task-adaptive design. Future work will focus on optimizing the model for mobile devices to provide real-time decision support in clinical settings.

Summary updated

Abstract

<title>Abstract</title> In-vitro fertilization (IVF) offers a solution for couples facing infertility challenges. However, the success of IVF, particularly in achieving live-birth outcomes, heavily depends on embryologists to conduct morphological assessments of fertilized embryos, a process that is both time-consuming and labor-intensive. While artificial intelligence (AI) has gained recognition for its potential to automate embryo selection, the application of deep learning (DL) is constrained by privacy concerns associated with the requirement for centralized training on extensive datasets. In this paper, we have developed a distributed DL system, termed ‘FedEmbryo’, tailored for personalized embryo selection while preserving data privacy. Within FedEmbryo, we introduce a Federated Task-Adaptive Learning (FTAL) approach with a hierarchical dynamic weighting adaption (HDWA) mechanism. This approach first uniquely integrates multi-task learning (MTL) with federated learning (FL) by proposing a unified multitask client architecture that consists of shared layers and task-specific layers to accommodate the single- and multi-task learning within each client. Furthermore, the HDWA mechanism mitigates the skewed model performance attributed to data heterogeneity from FTAL. It considers the learning feedback (loss ratios) from the tasks and clients, facilitating a dynamic balance to task attention and client aggregation. Finally, we refine FedEmbryo to address critical clinical scenarios in the IVF processes, including morphology evaluation and live-birth outcomes. We operate each morphological metric as an individual task within the client's model to perform FTAL in morphology evaluation and incorporate embryo images with corresponding clinical factors as multimodal inputs to predict live-birth outcomes. Experimental results indicate that FedEmbryo outperforms both locally trained models and state-of-the-art (SOTA) FL methods. Our research marks a significant advancement in the development of AI in IVF treatments.
BibTeX
@inproceedings{wang2024federated,
  author = {Guangyu Wang and Tianrun Gao and Yuning Yang and Kai Wang and Yuanxu Gao and Li-Shuang Ma and Lei Chen and Guangdong Liu and Ping Zhang and Xiaohong Liu},
  title = {{Federated Task-Adaptive Learning for Personalized Selection of Human IVF-derived Embryos}},
  booktitle = {Research Square},
  year = {2024},
  doi = {10.21203/rs.3.rs-4631058/v1},
}
Figure