Abstract
Abstract
Cell-type annotation in three-dimensional fluorescence microscopy often relies on dedicated molecular markers. Yet fluorescence channels are limited and often needed to label other structures of interest, so assigning one to a cell-type marker comes at a cost. A nuclear stain such as DAPI, by contrast, is present in nearly all protocols; classifying cell types from DAPI alone would therefore apply broadly to existing datasets while freeing channels for other structures. Here, we tested whether frozen self-supervised volumetric representations can support marker-sparing classification of hepatic 3D nuclear morphotypes from DAPI alone. We analyzed 14,385 nuclei from five adult mouse liver volumes and assigned marker-assisted reference morphotypes comprising hepatocyte, stellate, Kupffer, endothelial, and other cells. Using leave-one-animal-out cross-validation, we compared frozen 3DINO embeddings followed by validation-selected classical classifiers with 31 handcrafted DAPI-derived nuclear features and direct ResNet3D-18 classification. The 3DINO-based procedure achieved the highest mean held-out performance, with macro-F1 of 0.596 ± 0.041, balanced accuracy of 0.625 ± 0.037, and weighted F1 of 0.831 ± 0.046. ResNet3D-18 performed similarly (macro-F1 0.584 ± 0.051), whereas handcrafted features yielded lower mean performance (0.511 ± 0.040). Class-specific analysis showed near-perfect hepatocyte classification but greater ambiguity among non-parenchymal morphotypes. Thus, a frozen volumetric representation pretrained outside fluorescence microscopy achieved similar observed held-out performance to a task-specific 3D CNN trained end-to-end, while requiring only lightweight downstream classifier training. This provides a marker-sparing strategy for hepatic nuclear morphotyping while preserving fluorescence channels for other biological readouts.