A DEEP LEARNING FRAMEWORK FOR MULTI-OBJECT TRACKING AND PREDICTIVE TRAJECTORY ANALYSIS OF SPERM MOTILITY

Abstract

Male infertility is a significant global health challenge, and the accurate assessment of sperm motility is critical for reliable diagnosis and clinical intervention. Currently, Computer-Aided Semen Analysis (CASA) systems are the standard for quantifying motility, but they fundamentally struggle to maintain accurate tracking in physiologically relevant, high-density phase-contrast microscopy videos. Frequent cell crossovers, variable illumination, and occlusions cause identity switches and fragmented trajectories, which corrupt downstream kinematic parameters and restrict analyses to artificially short, non-physiological timeframes.

To address these limitations, this thesis proposes an integrated deep learning framework designed for robust, long-term, and identity-preserving sperm tracking in crowded environments. The pipeline combines a YOLO-based object detector for precise, centroid-accurate cell localization, with the ByteTrack multi-object tracking algorithm. ByteTrack's integration of predictive kinematics and low-confidence detection recovery is specifically leveraged to bridge temporary occlusions and preserve individual cell identities during complex crossover events. To support this supervised approach, a comprehensive training and evaluation pipeline was developed, encompassing manually curated real-world microscopy videos, rigorous data augmentations, and a stochastic agent-based synthetic simulator designed to faithfully replicate progressive cellular crowding.

Quantitative evaluation against classical heuristic tracking methods, specifically Trackpy, demonstrates the distinct superiority of the proposed framework across multiple tracking and biological cell-tracking metrics. The YOLO and ByteTrack pipeline achieved near-perfect accuracy in sparse conditions and maintained highly robust temporal associations in dense biological samples, significantly outperforming the baseline in Higher Order Tracking Accuracy (HOTA) and Identity F1 Score (IDF1). By successfully mitigating identity switches and extending stable trajectory reconstruction, this framework bridges the gap between short-term descriptive analysis and long-term predictive motility modeling, offering a substantial computational improvement for automated reproductive health assessments.