Automated Cytoplasmic Image Analysis System for Non-invasive Oocyte Quality Screening Workflow Using Static Bright-Field Images.

To determine whether a pre-trained deep neural network (ResNet50) can non-invasively evaluate oocyte quality and predict subsequent embryonic development by extracting cytoplasmic features from static bright-field images acquired immediately before piezo-intracytoplasmic sperm injection. This prospective cohort study included 306 metaphase II oocytes obtained from 63 patients (73 cycles). Immediately before piezo-intracytoplasmic sperm injection, spindle visibility was scored using PolScope, and static bright-field images were acquired. The cytoplasmic regions were automatically segmented and fed into a pre-trained ResNet50 model to extract intermediate layer features from layers 1 and 3. The resulting feature vectors were reduced to 20 components via principal component analysis. Logistic regression and support vector machines with stratified five-fold cross-validation were used to evaluate classification performance for spindle presence, normal fertilization, blastocyst formation, and high-quality blastocyst rates. Single ResNet-derived features were assessed individually. Logistic regression principal component analysis of 20 components achieved an area under the curve of 0.679 for spindle detection (sensitivity: 84.2%; specificity: 45.5%) and 0.521 for fertilization (sensitivity: 92.8%; specificity: 18.8%). Our support vector machine yielded marginal improvements. Single-feature analysis demonstrated high specificity for spindle detection (area under the curve, 0.656; specificity, 93.9%) and moderate specificity for fertilization (area under the curve, 0.607; specificity, 81.2%). Blastocyst prediction remained limited (area under the curve ≤ 0.64). Spindle-negative oocytes exhibited markedly better fertilization prediction (area under the curve, 0.835; specificity, 81.8%). Our deep neural network-based automated cytoplasmic image analysis system quantitatively and non-invasively assessed oocyte maturity without requiring large training datasets and predicted subsequent fertilization.