Table 3 Transformer applications in histopathological image detection and localization
Methods | Tissue | Dataset | Challenge | Highlight | ACC / F1 (%) |
|---|---|---|---|---|---|
GasHis-transformer [19] | Stomach (Gastric) | HE-GHI-DS | Inability of CNN models to handle global information well | GasHis-transformer and LW-GasHis-transformer | 97.97 / 97.97 |
PathTR [91] | Breast | CAMELYON16 | Neglecting the intrinsic WSI global correlations among the patches | Context-Aware Memory ViT with a CNN Backbone | 98.91 /– |
PVTCB-Lymph-Det [92] | Colon, breast and prostate | LYSTO | Detecting lymphocytes automatically due to the presence of artifacts and morphological variations | Pyramid ViT-based network and convolution attention mechanism with ResNet-50 | –/ 88.92 |
YOLOv5-transformer [93] | Breast, Colon, etc. | Custom | Accurate mitoses detection and morphological variations | Improved YOLOv5 transformer-based architecture | –/ 77.00 |
RAMST [94] | Stomach and colorectal | TCGA (CRC and STAD) | Unstable predictions caused by noisy patches and aggregation techniques | Joint regional attention and multi-scale transformer network | – |
CB-HVTNet [95] | Colorectal, breast, etc. | LYSTO and NuClick | Insufficient feature representations | Channel-boosted hybrid ViT network | –/ 80.00 |
Hossain et al. [96] | Breast, etc. | TCGA and Custom | ViT-based network | ROI selection ViT-based network | 96.10 /– |
PersAM [18] | Lymph | Custom | Attention region estimation in digital pathological images | Personalized attention mechanism ViT network | 83.13 /– |