EXNet: An Improved U-Net Architecture for Accurate Sperm Segmentation Through Spatial Feature Extractor and Multi-Scale Attention

Sperm morphology is crucial to male infertility diagnosis. Male infertility diagnosis requires better sperm morphology for assessing sperm quality. Recent advancements have significantly improved the ability to segment and evaluate sperm cells. Despite the improvements, a significant research gap remains in achieving accurate sperm segmentation, particularly in optimizing performance metrics. We propose an improved U-Net architecture for accurate sperm segmentation through spatial feature extractor and multi-scale attention “EXNet” to solve the segmentation challenges employed in human sperm segmentation. We employed the U-Net network architecture and a unique extractor and multi-scale attention mechanism. The attention mechanism extracts and refines the relevant spatial dimension features on multi-scale. The Extractor module aggregates information across multiple spatial locations in a feature map. A double convolutional down-sampling module and attention-based up-sampling extract and refine the spatial feature to improve segmentation for precise sperm component identification under difficult imaging circumstances. With limited computing power, our method outperforms the other U-Net methods in the separation of human sperm correctly. The model had a 95.14% DSC and a 65.63% MIOU following examination of sperm images. Performance and computational efficiency are balanced in EXNet. The findings show EXNet's reproductive health automated sperm analysis.