Computer Vision algorithms for microscopy images

Deep learning in computer vision has achieved remarkable results in visual recognition; however, its application to biological microscopy remains far more challenging than to natural images. This is due to low signal-to-noise ratios, uneven illumination, and dense, heterogeneous cellular environments that contain multiple similar organelles. Models capable of reliably interpreting these structures play a crucial role in diagnostics, as well as in genomic and molecular studies, where precise morphological characterization can support the identification and study of pathological conditions.\\ In this thesis, We investigate two complementary computer vision approaches applied to Transmission Electron Microscopy (TEM) images. First, we show that a standard, supervised U-Net architecture can achieve robust performance in mitochondrial semantic segmentation, highlighting that its main limitation lies not in the model itself, but in the scarcity of well-annotated data. We then explore a fully self-supervised alternative, training a ResNet-50 backbone with a self-supervised approach on images from patients affected by Gaucher disease (GD) ---a lysosomal storage disorder characterized by high phenotypic heterogeneity--- to learn general-purpose representations directly from unlabeled data. We show that the resulting trained backbone is able to capture meaningful morphological features and clinically relevant patterns without explicit supervision, demonstrating the ability not only to distinguish between different subcellular structures but also to capture the underlying signals that allow patient groups to be differentiated according to disease severity.