C-LiTE: A 22M-Parameter Foundation Model for Cellular Histopathology
C-LiTE is a lightweight foundation model for cellular representation learning in H&E-stained histopathology with 22 million parameters. It uses a native-resolution tokenisation strategy that processes 64 x 64 px cell-centred crops with 4 x 4 patch embedding into a 16 x 16 token grid without image upsampling. The authors adapt patch embedding weights from a pre-trained ViT-5-S via a pseudoinverse-based transformation and assemble a pre-training dataset of 60 million cell-centred crops from TCGA across 19 cancer types and 16 tissue types. On seven public datasets covering mitotic figure recognition, nucleus and cell-type classification, and immune cell phenotyping, C-LiTE reaches a mean macro-F1 of 0.670 under frozen-feature evaluation, compared with 0.682 for CytoFormer with roughly 30x fewer parameters.