Advancing whole slide image classification through pathology-pre-trained large language models

While multiple instance learning (MIL) shows potential for whole slide image (WSI) analysis in computational pathology, their clinical utility remains limited by the domain gap when using natural image-pretrained models, like ResNet-50, for feature extraction. Currently, pathology large language models (LLMs) show outstanding performance across various pathological tasks. However, few studies have compared the classification performance and generalization of ResNet-50 and pathology LLMs on different MIL methods. To provide a reference for the use of large-scale pre-trained models in computational pathology, and suggest promising research directions, in this paper, we systematically evaluate pathologyspecific LLMs against conventional convolutional architectures under four popular MIL methods. Our experiments reveal that pathology LLMs achieve higher accuracy than ResNet-50 across different MIL frameworks while exhibiting superior generalization in cross-domain validation. Additionally, LLMs trained by pathology images often obtain better performance than that trained by natural images and fine-tuned by pathology images. Moreover, the advantages of pathology LLMs are also influenced by different MIL methods.