Two-Stage Probabilistic Classification of Sperm Identity and Morphology

Background/Objectives: From a histological and embryological perspective, sperm identification and morphological classification are distinct assessment tasks. This study compares image-level error distributions and review-referral outcomes between separately trained conditional components and direct three-class learning. Methods: Following duplicate-image and label audits, 2947 images from the Sperm Morphology Image Dataset (SMIDS) were analyzed without changing the source labels. Logistic regression models using frozen EfficientNet-B0 features were evaluated using five outer and three inner cross-validation folds that preserved 2942 similarity groups. Final class probabilities, component probabilities, and outcomes at matched review volumes were compared. Results: Overall accuracy was 83.98% for the two-stage model and 84.80% for the direct model. Misclassification of non-sperm images as sperm decreased from 94 (direct) to 77 (two stage), whereas misclassification of true sperm images as non-sperm increased from 82 to 141. When each model referred 973 images for review, accuracy among the remaining 1974 images was 93.67% and 94.48%, respectively. Conclusions: The two models differed in the trade-off between erroneously including non-sperm material and excluding assessable sperm images. Overall and selective accuracy were higher for the direct model. The study provides a framework for jointly evaluating these two error directions on the same test images and at matched review volumes in preliminary histomorphological assessment.