MTCA-Net: Multi-Task Cascade Analysis Network for Real-Time Sperm Quality Analysis

Intracytoplasmic sperm injection (ICSI) critically relies on selecting high-quality sperm, which remains a labor-intensive and subjective process in clinical practice. Current deep learning algorithms involve trade-offs between accuracy and speed, resulting in inefficient real-time performance of sperm morphology and motility analysis. A multi-task cascaded analysis network (MTCA-Net) that integrates a detection module, a tracking module, and a segmentation module is proposed for real-time multi-dimensional sperm quality assessment. In the detection module, an attention mixed feature downsampler is proposed to improve the utilization of spatial down-sampling features. In the segmentation module, the proposed EffiFusion-U²Net achieves high precision with minimal parameters by incorporating weighted feature fusion blocks to dynamically adjust fusion weights. To enhance the model's generalization ability, an augmentation strategy based on FastGAN and pix2pixGAN was implemented to synthesize sperm heads. Additionally, three datasets are provided containing a total of 3,563 images in various resolutions for sperm detection and segmentation. Comprehensive evaluations on our datasets and four public datasets demonstrate that MTCA-Net achieves state-of-the-art performance in all cases. MTCA-Net is capable of performing high-precision segmentation at 51 frames per second (FPS) and stable tracking at 94 FPS on high-resolution images of 1920 × 1200 pixels, meeting the real-time analysis requirements of clinical ICSI.