Multiple instance learning-based prediction of programmed death-ligand 1 (PD-L1) expression from hematoxylin and eosin (H&E)-stained histopathological images in breast cancer.

Programmed death-ligand 1 (PD-L1) is an important biomarker increasingly used as a predictive marker in breast cancer immunotherapy. Immunohistochemical quantification remains the standard method for assessment. However, it presents challenges related to time, cost, and reliability. Hematoxylin and eosin (H&E) staining is a routine method in cancer pathology, known for its accessibility and consistently reliability. Deep learning has shown the potential in predicting biomarkers in cancer histopathology. This study employs a weakly supervised multiple instance learning (MIL) approach to predict PD-L1 expression from H&E-stained images using deep learning techniques. In the internal test set, the TransMIL method achieved an area under the curve (AUC) of 0.833, and in an independent external test set, it achieved an AUC of 0.799. Additionally, since RNA sequencing results indicate a threshold that allows for the separation of H&E pathology images, we further validated our approach using the public TCGA-TNBC dataset, achieving an AUC of 0.721. These findings demonstrates that the Transformer-based TransMIL model can effectively capture highly heterogeneous features within the MIL framework, exhibiting strong cross-center generalization capabilities. Our study highlights that appropriate deep learning techniques can enable effective PD-L1 prediction even with limited data, and across diverse regions and centers. This not only underscores the significant potential of deep learning in pathological artificial intelligence (AI) but also provides valuable insights for the rational and efficient allocation of medical resources.