Thesis Defense: Mostofa Uddin | November 18, 2025 | 1:00pm

Committee:

The cell exerts its functions through individual and combined mechanisms of numerous subcellular objects-from large organelles to small macromolecules. The morphology of these objects plays a crucial role in the determination of cellular behavior and responses to perturbations. The advent of in situ high-resolution microscopic imaging technologies has enabled the observation of these structures within their native cellular environments. However, a substantial algorithmic gap remains in accurately detecting and characterizing subcellular objects from these images. Existing methods rely either on annotation-dependent supervised learning or non-learning approaches with poor scalability and accuracy, making them ineffective at detecting subcellular objects with rare or unusual morphologies. This thesis addresses these challenges by developing a series of novel unsupervised algorithms for the detection and morphological characterization of subcellular objects from microscopy images across multiple spatial scales. First, we introduce a multiscale segmentation method that leverages the implicit pattern-recognition capabilities of vision foundation models to segment subcellular objects of varying sizes from noisy microscopy data. To further specifically address the localization of small macromolecules in 3D cellular microscopy, we propose a positive-unlabeled learning based voxel classification approach. While segmentation often suffices for characterization of large objects, the same is not true for objects that are small relative to the image scale. To identify the morphology of small objects, we developed an unsupervised disentangled representation learning method that disentangles transformations from the intrinsic semantic content of the images. Building on this, we established pipelines for two major applications: (1) investigating subtype-specific nuclear morphologies throughout aging in mouse and human neurons by tailoring the method for RNAscope images and (2) identifying in situ morphologies of macromolecules from cellular cryo-electron tomography images, solving a largely unresolved problem, by extending the method for SE(3) disentanglement. Finally, we develop a novel implicit-disentanglement method that directly identifies different compositions and conformations of macromolecules from these images. Together, these methods enable robust and automated detection and characterization of subcellular objects across different microscopy imaging modalities and spatial scales, reveal previously unknown morphologies, and advance our understanding of the subcellular world.