Refined trajectory smoothing and deep learning classification of human sperm motility

Abstract STUDY QUESTION Can precise trajectory smoothing improve extraction of sperm motility features, and can deep learning on raw trajectory data enable accurate classification of sperm motility patterns? SUMMARY ANSWER We present an approach that enhances the precision of motility parameter extraction through frequency-domain smoothing and enables accurate classification of sperm motility patterns using a deep learning model trained on raw trajectory data. WHAT IS KNOWN ALREADY Conventional computer-aided sperm analysis (CASA) systems estimate motility parameters by applying basic smoothing algorithms to derive an average path, which can result in over- or under-smoothing, leading to inaccuracies in key parameters such as beat cross frequency (BCF) and amplitude of lateral head displacement (ALH). Since the identification of hyperactivated spermatozoa relies heavily on these kinematic metrics, such inaccuracies can contribute to misclassification. STUDY DESIGN, SIZE, DURATION This cross-sectional study analysed 2326 sperm trajectories (1931 progressive, 395 hyperactivated) recorded at 60 frames per second, derived from five individual samples, to develop and evaluate improved motility parameter extraction methods and trajectory-based classification models. PARTICIPANTS/MATERIALS, SETTING, METHODS We compared Gaussian Process Regression (GPR), moving average, and Discrete Cosine Transform (DCT) smoothing to improve average path estimation. A novel metric, path average width (PAW), was introduced to quantify lateral head displacement. An ensemble of InceptionTime models was trained on (x, y) coordinate sequences to classify spermatozoa as progressive or hyperactivated. Additional classification of motility grades was performed using trajectory endpoints. MAIN RESULTS AND THE ROLE OF CHANCE The DCT model retaining 12 frequency components (DCT-12) produced the most consistent and symmetric average paths, leading to improved accuracy in the calculation of BCF and ALH. Our introduced PAW metric effectively distinguished between hyperactivated spermatozoa (5.5 ± 1.5 μm) and progressive spermatozoa (2.0 ± 1.3 μm). The InceptionTime-based classification model achieved 89% accuracy in differentiating progressive and hyperactivated trajectories, and 78% accuracy for predicting motility grades. LIMITATIONS, REASONS FOR CAUTION Models were trained on sperm trajectories recorded in low-viscosity media. Since sperm selection for ICSI is performed in viscous environments like low concentrations of polyvinylpyrrolidone, future training on such data is essential to improve clinical translation. Additionally, the model for classifying progressive and hyperactivated sperm was trained on a single-centre dataset (5 individuals, total of 790 trajectories) and, despite cross-validation and data augmentation, still requires independent, multi-centre validation to confirm generalizability. Absence of personal identifiers and clinical metadata precluded per-person analyses. WIDER IMPLICATIONS OF THE FINDINGS By integrating refined signal-based feature extraction with trajectory-level classification, our method addresses core limitations of CASA systems and holds potential for real-time application into ART workflows. Training on high-viscosity media could further enhance its applicability to sperm selection for ICSI. STUDY FUNDING/COMPETING INTEREST(S) This work was supported by the Australian Research Council (ARC) Discovery Project Grants (DP210103361 to A.N. and R.N.), the Australian National Health and Medical Research Council (NHMRC) fellowship (Investigator Grant 2017370 to R.N.), and Monash IVF Group support. The authors declare no competing interests. TRIAL REGISTRATION NUMBER N/A.