ADVANCING SPERMIOGENESIS ANALYSIS: MACHINE LEARNING-BASED APPROACH TO OBJECTIVE CLASSIFICATION OF SPERMATID STEPS: Congress of the American Society of Andrology & International Congress of Andrology 2025

Introduction and Objective: Spermiogenesis is a dynamic process
essential for sperm development. During this process, round spermatids
undergo modifications that allow them to differentiate into
elongated spermatids, ultimately leading to mature spermatozoa.
These changes include acrosome formation, sperm head elongation,
flagellum development, and chromatin remodeling. Spermatid classification
usually relies on expert visual criteria, based on the topological
morphology of cells as observed in histological preparations
of testes. However, there is a need for more advanced and objective
techniques to differentiate stages in isolated mixed germ cell
populations.
Methods: We developed a method to objectively classify each spermatid
step using a machine-learning approach. Germ cells from mouse
testes (n = 5) were isolated using enzymatic digestion and labeled with
DAPI as a nuclear marker, PNA-FITC as an acrosome marker, and antiacetylated
tubulin as a manchette marker. Fluorescence microscopy
was used to image each cell and its respective fluorescent labels, while
phase contrast microscopy was employed for cytoplasmic assessment.
Next, morphological analyses were performed on each structure in
n = 405 individual cells using Nuclear Morphology Analysis software
v1.20.0, and six variables (area, perimeter, maximum Feret, minimum
diameter, bounding height, and bounding width) were measured.
Then, mathematical algorithms, including principal components analysis
(PCA), as well as unsupervised (k-means clustering) and supervised
(linear discriminant analysis [LDA]) machine learning approaches,were
applied using GraphPad Prism Pro software.
Results: Based on the Kaiser criterion and an eigenvalue >1, the PCA
identified two principal components summarizing the most important
features among all the variables analyzed. Moreover, the k-means
clustering algorithm classified the dataset into five subpopulations
according to morphological criteria, differentiating the spermatids into
step 8, steps 9–10, steps 11–12, steps 13–14, and steps 15–16, with a silhouette score of 0.5. Next, LDA analysis validated these results,
revealing that 97.2% of the cells were classified correctly,with an error
rate as low as 2.8%.
Conclusion: This study developed an objective, machine-learningbased
method to classify spermatid steps with high accuracy. The PCA
reduced the complexity of the dataset to two key PCs, while k-means
clustering classified the spermatids into distinct steps. LDA further validated
these classifications with a high accuracy rate. This approach
provides a reliable tool for spermatid classification, which can enhance
our understanding of the sperm differentiation process and facilitate
future research in reproductive biology.