Human embryo stage classification using an enhanced R(2 + 1)D model and dynamic programming with optimized datasets

Abstract

Infertility affects millions of couples worldwide, with Assisted Reproductive Technology (ART), particularly In Vitro Fertilization (IVF), offering hope for many. The success of IVF critically depends on accurately assessing embryo quality. Traditional assessment methods relying on subjective morphological criteria face significant limitations, underscoring the need for more reliable approaches. This study proposes an advanced model for embryo evaluation, leveraging a three-dimensional deep learning framework based on a refined ResNet-R(2 + 1)D architecture. The model incorporates Spatial-only Self-Attention (SSA) and Squeeze-and-Excitation (SE) blocks to enhance spatial and channel-wise feature extraction. Additionally, convolutional blocks (convB) are integrated before and after the network to align feature representations effectively. Dynamic programming with the Viterbi algorithm ensures biologically consistent predictions during post-processing. The model is trained on a balanced and meticulously pre-processed dataset of time-lapse microscopy images, addressing issues of data imbalance and quality. Experimental results demonstrate the proposed model’s exceptional performance, achieving an accuracy of 93.3 %—a 13.1 % improvement over the baseline R(2 + 1)D model trained on a balanced dataset. Compared to state-of-the-art methods, the proposed model demonstrates acceptable accuracy and scalability, effectively managing the classification of 15 embryo developmental stages. These findings highlight the significant potential of advanced deep learning techniques in improving embryo selection and enhancing IVF success rates.