Answers are grounded in this paper and cite their sources.
The success of in vitro fertilization (IVF) depends on the manual skill of embryologists, which introduces operator variability and limits standardization across laboratories. Automating IVF laboratory procedures could improve consistency. A prerequisite for such automation is the reliable detection and segmentation of gametes and instruments in microscopy images. In this single-center feasibility study, we fine-tuned off-the-shelf deep learning models by transfer learning and evaluated them on three computer vision tasks in intracytoplasmic sperm injection (ICSI): detection of the oocyte and instrument tips (RT-DETR-L), detection of the sperm inside the injection needle (RT-DETR-L), and instance segmentation of the sperm head and tail (YOLO11-L-seg). Images were extracted from videos of clinical ICSI procedures recorded with a custom inverted microscope, and the data were partitioned at the patient level, so that no patient contributed to more than one subset. On the held-out test sets, the three models reached an mAP@0.50 of 0.988, 0.962, and 0.990, respectively, with model inference latencies of 7.0–11.6 ms per image on a single GPU. These results suggest that off-the-shelf models, adapted with modest clinical datasets, are a feasible starting point for the perception layer of robotic ICSI.
No takes yet. Share an insight, caveat, or question.
Reyes-Llamas et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: