Generalizable Learning for 2D and 3D Visual Understanding From General to Biomedical Imaging

Recent advances in deep learning have significantly improved visual understanding, enabling machines to achieve human-level performance in recognition, detection, and segmentation. However, while progress in natural image domains has been remarkable, biomedical imaging continues to face challenges due to limited annotations, domain shifts across acquisition conditions, and the inherent complexity of multiview and three-dimensional data. These limitations underscore the need for generalizable learning frameworks that can adapt across diverse tasks and modalities. This thesis explores strategies to enhance generalization in both 2D and 3D visual understanding, bridging general computer vision and biomedical applications. We investigate zero-shot learning to extend recognition beyond pre-defined categories, domain adaptation to mitigate distributional gaps between simulated and experimental data, and multi-view and graph-based methods to better model complex structural information. The thesis contributes five major research advances: (1) the first zero-shot framework for camouflaged object detection in natural images, (2) a new stem cell dataset with a robust segmentation pipeline for astrocyte imaging, (3) a multi-view transformer (MVGFormer) tailored for cryo-electron tomography segmentation, (4) voxel-wise unsupervised and zero-shot domain adaptation frameworks for cryo-ET subtomogram segmentation, and (5) a multi-view graph fusion framework for diabetic retinopathy grading from fundus images. Together, these contributions establish a coherent pathway toward robust and generalizable learning across visual domains. The proposed methods not only advance the state of the art in each individual task but also provide a foundation for future biomedical foundation models that can integrate multi-modal and multi-scale data for clinical translation.