AI-assisted diagnostics in male infertility: morphological assessment, molecular stratification, and the translational gap between laboratory performance and clinical utility

Background Male infertility contributes to approximately half of all cases of couple infertility worldwide, yet its diagnostic cornerstone, conventional semen analysis anchored in World Health Organization (WHO) criteria, suffers from inter-observer variability that can reach 30–40% for morphological parameters even among trained andrologists. Computer-assisted sperm analysis (CASA) reduced some sources of error but failed to eliminate the subjectivity inherent in morphological classification, particularly for the acrosome, midpiece, and cytoplasmic droplet defect categories. Convolutional neural networks and ensemble deep learning architectures have since demonstrated high classification accuracy on selected benchmark datasets (detailed under Results), prompting fundamental questions about the clinical readiness and translational utility of AI-assisted andrology. Methods (Approach) This is a critical review that synthesises evidence from peer-reviewed studies identified through targeted searches of PubMed, Scopus, and IEEE Xplore (through May 2025), covering AI applications in sperm morphology classification, motility analysis, non-obstructive azoospermia (NOA) outcome prediction, and sperm selection for intracytoplasmic sperm injection (ICSI), supplemented by hand-searching the reference lists of key articles. Search terms combined controlled vocabulary and free-text terms for “sperm morphology,” “deep learning,” “convolutional neural network,” “azoospermia,” “testicular sperm extraction,” and “ICSI,” restricted to English-language, peer-reviewed publications. Studies were selected for methodological relevance, recency, and clinical significance rather than through a pre-registered, reproducible systematic protocol; no formal risk-of-bias or quality-appraisal instrument (e.g., QUADAS-2 or PROBAST) was applied. This is a deliberate scope decision, declared here rather than implied: the review is a critical, evidence-synthesising appraisal of a fast-moving field, not a systematic or scoping review, and quantitative performance figures drawn from heterogeneous studies (summarised in Table 1) should be read as illustrative of the state of the field rather than as directly comparable, pooled estimates. Results Deep learning models, including ensemble and attention-augmented convolutional architectures, have reported accuracies of 94% to 96.77% and F1-scores of up to 98.2% for sperm morphology classification, principally on the HuSHeM benchmark dataset; the morphology-classification studies summarised in Table 1 report accuracy, precision, or F1-score rather than AUC (area under the receiver operating characteristic curve). Machine learning models integrating hormonal, genetic, and histopathological features achieve AUC values of 0.807 to above 0.90 for predicting sperm retrieval success in microdissection testicular sperm extraction (m-TESE). For AI-assisted real-time sperm selection during ICSI, a prospective sibling-oocyte study of 646 metaphase II oocytes found only a non-significant trend towards improved fertilisation, whereas a separate, smaller sibling-oocyte comparison, reported to date only as a conference abstract, described a statistically significant improvement in fertilisation; live birth rate data remain scarce. However, the predominance of single-centre, algorithmically homogeneous training datasets, absence of prospective multicentre validation, and the opaque decision logic of most deployed models collectively constrain clinical adoption. Conclusions AI-assisted sperm morphology assessment has attained technical maturity but not clinical readiness. The transition from benchmark accuracy to actionable diagnostic utility requires prospective multicentre validation using clinically diverse and ethnically representative cohorts, live birth rate pre-specified as the primary endpoint, with fertilisation and blastulation rates as secondary endpoints, and the development of explainable AI frameworks compatible with regulatory standards including the EU AI Act. The Middle East and North Africa region, characterised by high rates of male-factor infertility, distinct genetic aetiologies, and constrained laboratory infrastructure, represents both a critical validation environment and an equity-relevant deployment target for these technologies.