Deep Learning-Based Imaging Informatics for Automated Detection of Cytoplasmic Strings in Embryo Time-Lapse Microscopy.

Cytoplasmic strings (CS) are difficult to identify in embryo time-lapse microscopy because they are thin, low-contrast, and variably oriented. This retrospective single-center study developed and evaluated an oriented object-detection framework for CS localization. Three embryologists independently annotated 3066 frames from 171 blastocyst videos using oriented bounding boxes (OBBs), and data were partitioned at the video level. YOLOv8m, YOLO11m, and YOLO12m were evaluated with and without a Convolutional Block Attention Module (CBAM) using rotated intersection-over-union, instance-level center localization, and 613 unanimously confirmed CS-negative frames. Agreement was substantial for CS count categories (Fleiss' κ = 0.717 ) and good for box counts (ICC A,1 = 0.818 ). YOLO11m+CBAM achieved the highest observed mAP 50 on the positive-only and mixed sets (74.88% and 69.36%, respectively), whereas YOLO11m had higher strict mAP 50 - 95 and baseline models produced fewer negative-frame false positives. In a controlled comparison, OBB supervision reduced background enclosure and yielded higher observed performance than an axis-aligned baseline in a shared evaluation space. Within this retrospective single-center dataset, the evaluated models achieved measurable performance for cytoplasmic string instance localization. However, multicenter, temporal, frame-level, and outcome-linked validation is required before clinical use.