Abstract
The implantation potential of an embryo is intricately linked to the quality of its blastocyst. Consequently, achieving an objective and precise identification of blastocyst morphology is imperative. The purpose of this study is to focus on the structural associations between all tissues of blastocysts, and explore the possibility of semi-supervised learning (SSL) for blastocyst segmentation, so as to further improve the segmentation performance of blastocyst tissues. In this paper, we present a framework named I2C2Net for the automatic segmentation of blastocysts in human embryo images, leveraging both supervised and semi-supervised learning approaches. I2C2Net consists of three key components: the Intra-Class Context Module (IACCM), the Inter-Class Context Module (IRCCM), and the Consistency Module (CM). The IACCM aggregates pixel representations within specific category areas, learning categorized regions relative to ground truth labels. This aggregation then decomposes a K-category recognition task into two tasks, each with distinct labels, while retaining the ability to learn intra-class features. The design of the IRCCM is informed by blastocyst morphology, capturing inter-class information on blastocyst tissues as they develop from inner to outer layers. Additionally, we introduce a consistency module for supervised training, enhancing the model's ability to learn the original data distribution and improve recognition accuracy without increased computational burden. Furthermore, to address the scarcity of annotated data and meet clinical demands, we propose a semi-supervised version of I2C2Net based on the semantic consistency assumption and cluster assumption across heterogeneous domains. The ablation experimental results validate the effectiveness of our proposed IACCM, IRCCM, and CM modules. Compared to other supervised methods, our I2C2Net achieves state-of-the-art performance in terms of Accuracy, Precision, Recall, Dice coefficient, and Jaccard index, which are [Formula: see text], [Formula: see text], [Formula: see text], [Formula: see text], [Formula: see text], respectively. Moreover, our semi-supervised version of I2C2Net achieves the best performance among other popular SSL approaches. Specifically, our model gains at least a [Formula: see text] rise in Accuracy, [Formula: see text] rise in Precision, [Formula: see text] rise in Dice coefficient, and [Formula: see text] rise in Jaccard index. The quantitative and qualitative experimental results showcase the superiority of our model over other representative supervised and semi-supervised methods on the blastocyst public dataset.