Table 4:
Overview of transfer learning models. Note: (✓) indicates the code is publicly available and the link is provided in their respective paper.
| Reference | Cancer types | Staining | Application | Method | Dataset |
|---|---|---|---|---|---|
| Wang et al. (2016a) | Breast | H&E | Detection of breast cancer metastasis | Pre-trained GoogleNet model | Camelyon16 (400 WSIs) |
| Liu et al. (2017) | Breast | H&E | Detection of breast cancer metastasis | Pre-trained Inception-V3 model | Camelyon16 (400 WSIs) |
| Han et al. (2017) | Breast | H&E | Breast cancer multi-classification | CNN integrated with feature space distance constraints for identifying feature space similarities | BreaKHis (7,909 images) |
| Lee and Paeng (2018) | Breast | H&E | Detection and pN-stage classification of breast cancer metastasis | Patch based CNN for metastasis detection + Random forest classifier for lymph node classification | Camelyon17 (1,000 WSIs) |
| Chennamsetty et al. (2018) | Breast | H&E | Breast cancer classification | Ensemble of three pre-trained CNNs + aggregation using majority voting | BACH 2018 challenge (400 WSIs) |
| Kwok (2018) | Breast | H&E | Breast cancer classification | Inception-Resnet-V2 based patch classifier | BACH 2018 challenge (400 WSIs) |
| Bychkov et al. (2018) | Colon | H&E | Outcome prediction of colorectal cancer | A 3-layer LSTM + VGG-16 pre-trained features to predict colorectal cancer outcome | Private set (420 cases) |
| Arvaniti et al. (2018) (✓) | Prostate | H&E | Predicting Gleason score | Pre-trained MobileNet architecture | Private set (886 cases) |
| Coudray et al. (2018) (✓) | Lung | H&E | Genomics prediction from pathology images | Patch based Inception-V3 model | TCGA (1,634 WSIs) and validated on independent private set containing frozen sections (98 slides), FFPE sections (140 slides) and lung biopsies (102 slides) |
| Kather et al. (2019) (✓) | Colon | H&E | Survival prediction of colorectal cancer | Pre-trained VGG-19 based patch classifier | TCGA (862 WSIs) and two other public datasets (25 + 86 WSIs) |
| Noorbakhsh et al. (2019) (✓) | Multi-Cancers | H&E | Pan-cancer classification | Pre-trained Inception-V3 model | TCGA (27,815 WSIs) |
| Tabibu et al. (2019) (✓) | Kidney | H&E | Classification of Renal Cell Carcinoma subtypes and survival prediction | Pre-trained ResNet based patch classifier | TCGA (2,093 WSIs) |
| Akbar et al. (2019) | Breast | H&E | tumour cellularity (TC) scoring | Two separate InceptionNets: one for classification (healthy vs. cancerous tissue) and the other outputs regression scores for TC | BreastPathQ (96 WSIs) |
| Valkonen et al. (2019) (✓) | Breast | ER, PR, Ki-67 | Cell detection | Fine-tuning partially pre-trained CNN network | DigitalPanCK (152 - invasive breast cancer images) |
| Ström et al. (2020) | Prostate | H&E | Grading of prostate cancer | Ensembles of two pre-trained Inception-V3 models | Private set (8730 WSI’s) |