Revisiting the Role of Foundation Models in Cell-Level Histopathological Image Analysis under Small-Patch Constraints -- Effects of Training Data Scale and Blur Perturbations on CNNs and Vision Transformers
该研究通过系统评估发现,在细胞级(40x40 像素)小图块病理图像分析中,当训练数据规模充足时,针对小图块优化的任务特定架构(如 CustomViT)在准确性和效率上均优于基础模型,且基础模型并未展现出更强的模糊鲁棒性。