Detection of human sperm cells using deep learning-based object detection methods

Infertility has become a significant health issue worldwide in the last 50 years. This issue, with varying rates across different regions of the world, affects approximately one out of every ten couples on average globally. Diagnosis of male-related infertility is conducted by evaluating sperm quality. When investigating sperm quality, factors such as sperm count, motility, and morphological structure are assessed. Detection of sperm before the analysis of sperm motility and count is an important step. In this study, autonomous sperm detection was carried out using deep learning methods, namely Faster R-CNN and YOLOv3, on a newly generated and unique semen video dataset. This distinct dataset, created within the scope of this study, includes semen videos from 10 patients obtained with the assistance of a mobile phone under a microscope. Videos contain label information that classifies objects as sperm and non-sperm. Labeled videos prepared for analysis were evaluated under two scenarios: patient-focused and patient- independent. In the first scenario, eight labeled videos were combined to train and test Faster R-CNN and YOLOv3 models in three different ratios. In the second scenario, each trained model was tested with two videos that had never been part of the training process. In this second scenario, detection performances were evaluated using videos that had not been involved in training. The study achieved sperm detection results of approximately 96% in individual videos using the YOLOv3 model and an average mAP of 84.5%. When compared against two significant criteria, object detection accuracy and training times, the YOLOv3 method was observed to be more successful than the Faster RCNN method.