Development of Automated and Machine Learning-based image analysis for morphological characterization of cellular structures

Advanced microscopy techniques are increasingly used to study complex human in vitro models, including both two dimensional and three dimensional systems. These models provide biologically relevant information, but they also generate large and heterogeneous imaging datasets that are difficult to analyze through manual inspection alone. Manual quantification is also difficult to standardize, as the results may vary between operators and the analysis becomes impractical when applied to large image datasets. In this context, the aim of this thesis is to develop automated computational workflows for the extraction of quantitative information from microscopy images, combining segmentation, feature extraction and Machine Learning based classification. The work is applied to two experimental contexts: somatic cell reprogramming toward naïve human induced pluripotent stem cells (hiPSCs) and the analysis of synaptic alterations in forebrain organoids derived from control and Fragile X Syndrome (FXS) patient lines. In the first case, to characterize the reprogramming process, a hierarchical Machine Learning workflow is developed to process heterogeneous microfluidic cultures. After nuclear segmentation, quantitative nuclear morphological features and fluorescence measurements from specific population markers are extracted and used to classify target reprogrammed cells while separating them from feeder cells, debris and background artifacts. A second classification level is then used to distinguish different biological populations within the reprogrammed culture. In the long term, this could help to describe reprogrammed cultures using mainly morphological information, reducing the need to rely on specific population markers. In the second case, to study FXS patient-specific alterations, an automated workflow based on colocalization was optimized to quantify pre- and post-synaptic puncta in forebrain organoids at day 90. The analysis included organoids derived from three FXS patient lines, with one line analyzed as two independent experimental replicates, resulting in four FXS groups in the final quantitative comparison. The pipeline allows the standardized measurement of VAMP2, PSD95 and colocalized synaptic structures, normalized to the dendritic area. The analysis shows a reduction of colocalized synaptic puncta in two of the three FXS patient-derived lines compared with controls, supporting the investigation of synaptic phenotypic alterations associated with FXS. Overall, this thesis shows that automated image analysis and Machine Learning based approaches can improve the extraction of quantitative information from complex microscopy datasets. By reducing manual intervention and increasing standardization, these workflows provide useful tools for analyzing heterogeneous cellular systems in both 2D and 3D in vitro models.