Flow chart of three training strategies to generalize the virtual-staining model pre-trained with A549 and HEK293T cells to the BJ-5ta cell type with limited training samples. The bounding boxes indicate strategies: blue, virtual-staining pre-training from scratch with paired images of BJ-5ta; orange, pre-training with paired images of HEK293T and A549, and fine-tuning with paired images of BJ-5ta; green, FCMAE pre-training with only the phase images of HEK293T and A549, virtual-staining pre-training with images of HEK293T and A549, and fine-tuning with paired images of BJ-5ta. The pre-training steps initialize model weights in the encoder (FCMAE) and decoder (virtual staining) of UNeXt2.
