De Clemente, Claudia (2024) SINGLE-CELL TRACKING ANALYSIS FOR LABEL FREE SEMEN DIAGNOSIS BY MEANS OF NEURAL NETWORKS. [Tesi di dottorato]

	Documento PDF Claudia_De_Clemente_PhD_Thesis.pdf Download (7MB)
	Documento PDF De Clemente_Claudia_37_PARZIALE.pdf Visibile a [TBR] Amministratori dell'archivio Download (3MB) | Richiedi una copia

Abstract

Single-cell analysis allows the retrieval of precious information from a pool of cells that would not be appreciated with a bulk analysis. In fact, the characterisation of cell biophysical properties -such as area, perimeter, axes, density, stiffness, and electric potential- has proven to be a powerful mean for phenotyping, sorting, detection, drug testing applications, and to gain more insight into physiological cell functions in responses to external stimuli. Besides biophysical properties, single-cell tracking analysis is a powerful tool for high-precision diagnostics, gathering information on cell kinematics and morphology. Object tracking, which is the process of locating and following a specific object and its behaviour in sequential images, is a difficult task in biology because of the large size of cell image sequences, which makes the analysis more convenient to automatize. Indeed, computational object tracking can improve reproducibility and computational time significantly. However, the very small cell sizes often require fluorescent labels to ease the location and the tracking process, reducing the sample viability or potentially altering cell properties. Semen infertility diagnosis is based on the evaluation of a comprehensive set of sperm cells quality indicators, including motion and morphology. For this reason, it is the optimal context for developing innovative single-cell analysis approaches. The World Health Organization (WHO) establishes stringent criteria for assessing semen health. Presently, the gold standard for evaluating semen quality relies on visual examination of motility and morphological parameters, introducing limitations in accuracy and repeatability. However, in 1980, computer-assisted sperm analysis (CASA) systems emerged trying to reduce these analysis biases. In detail, CASA measures the motion of live spermatozoa through sperm head centroid tracking at low magnification, while -for morphology- it requires higher magnification as well as fixing and staining procedures, possibly compromising sample viability and integrity. Additionally, the inability to perform a simultaneous morphological and kinematic assessment on a single cell hinders precise sperm characterization. Addressing these limitations, various label-free techniques have been devised for both motility and morphology assessments, offering promising alternatives in semen quality evaluation. On this line, the integration of computer science, machine learning (ML) and data science allows the objective analysis of large amounts of data with high accuracy. In neural-network-based ML applications, the aim is the creation of an algorithm to train a machine to solve problems similarly to the human mind, capturing complex relationship between input and output data. For example, convolutional neural networks (CNNs) have been adopted for sperm motility categorization as well as sperm tracking. Deep convolutional neural networks (DCNNs) instead are more commonly used for single-sperm morphology evaluation, DNA fragmentation index retrieval and abnormalities identification. These quality metrics may be further evaluated through advanced ML algorithms, which are able to handle large amounts of data finding possible correlations between sperm parameters at the single-sperm level. This further reduces analysis subjectivity and enhances diagnostic potential. The objective of the thesis is the development of an automated approach for label-free analysis of sperm cells, based on the automatic tracking of single cells to extract kinematic and biophysical features related to morphology and movement dynamics. Being label-free, the method preserves the viability of the sample, making the process fast and easily reproducible in a clinical context. The procedure involves capturing a brightfield image sequence of swimming sperm in a glass capillary at 6 40x magnification, followed by single-sperm detection, tracking and segmentation. The detection and segmentation steps are performed with two neural networks, YOLO4-tiny and U-net tiny respectively. Subsequently, kinematic and morphometric parameters are evaluated. Of note, three parameters are introduced as novel features to better characterize sperm motion. To verify the neural network generalization ability and the new parameters to detect sperm heterogeneity, the analysis was performed under several conditions resembling physio-pathological contexts. In particular, variations in viscosity, osmolarity, absence of calcium, and high doses of caffeine were adopted. This allowed the retrieval of a final dataset of 3900 cells. Principal Component Analysis (PCA) and clustering with the K-means algorithm were adopted to characterize sperm populations in more detail compared to the standard classifications found in the literature. Four clusters were identified, identifying different grades of progressivity related to both motility and head morphology. The outcome demonstrate that the proposed approach allows for advanced characterization of sperm populations, providing more information on motility compared to traditional CASA systems. Furthermore, the employed cascade of convolutional neural networks provided a more accurate and comprehensive description of sperm movement in heterogeneous conditions.

Downloads

Actions (login required)

Modifica documento