Table 5 Transformer applications in histopathological image representation
Method | Tissue | Dataset | Challenge | Highlight | ACC / AUC (%) |
|---|---|---|---|---|---|
CD-Net [105] | Breast, Lung | TCGA (LUAD and LUSC) | Inability to leverage the rich multi-resolution information | Transformer-based pyramidal context-detail network | 91.10 / 95.80 |
DSCAÂ [106] | Lung, breast and brain | NLST, TCGA (BRCA and LGG) | High computational complexity and unnoticed semantic gap in multi-resolution feature fusion | A dual-stream Transformer network with cross-attention framework | - |
HIPT [22] | Breast, lung, stomach, cell | IDC, LUAD, CCRCC, PRCC, CHRCC, and STAD | The structure of phenotypes in tumors and learning a good representation of a WSI | Hierarchical image pyramid Transformer with two levels of self-supervised learning | –/ 98.00 |
HEATÂ [107] | Colon, breast and esophageal | CAMELYON16, TCGA (COAD, BRCA, and ESCA) | The challenges of extracting diverse interactions between various cell types | Heterogeneous-graph edge attribute Transformer-based network | 99.90 / 99.90 |
H2TÂ [21] | Lung, breast and kidney | TCGA-NSCLC, CPTAC-LUAD, etc., BRCA, RCC and ACDC | High discordance on how a tissue sample and higher predictive power that comes at the cost of interpretability | Handcrafted histological Transformer-based network for unsupervised representation WSIs | - |
ViT-AMCNet  [20] | Laryngeal, breast, brain | Laryngeal cancer, breast cancer, brain cancer | Problems of poor transformer generalization bias and poor AMC interpretive ability | ViT-based network with adaptive model fusion and multi-objective optimization | 95.14 / 96.17 |