BTFormer: Blast transformer for human blastocyst components segmentation

Abstract

Assessing embryo quality through segmentation of blastocyst components is crucial, as embryo morphology directly correlates with its potential for implantation. However, automatic blastocyst segmentation remains a challenging task due to factors such as poor contrast, noise, and ambiguous boundaries between different tissue structures. In this study, we introduce a novel transformer-based architecture, termed BTFormer (Blastocyst Transformer), designed to effectively segment blastocyst components. Firstly, we use an axial-free attention mechanism with lower computational resources, which catches non-local feature maps with long-range cues to alleviate the mistake of local structure. Secondly, to enjoy the rotation consistency of the embryo images, we propose an axial-free attention block with a soft aggregation operation to embed features extracted by axial-free attention with different angles, which collect global cues and broadcast a diversified receptive field. We validated our method on a typical public dataset and achieved the state-of-the-art segmentation performance with accuracy, precision, recall, Dice coefficient, and Jaccard index of 93.86%, 91.81%, 92.25%, 92.02% and 85.45%. Extensive qualitative experimental results demonstrate the effectiveness of our proposed method.