Table 2:
Overview of weakly supervised 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 |
|---|---|---|---|---|---|
| Multiple instance learning (MIL) | |||||
| Hou et al. (2015) | Brain | H&E | Glioma subtype classification | Expectation-maximization based MIL with CNN + logistic regression | TCGA (1,064 slides) |
| Jia et al. (2017) | Colon | H&E | Segmentation of cancerous regions | FCN based MIL + deep supervision and area constraints | Two private sets containing colon cancer images (910+60 images) |
| Liang et al. (2018) | Stomach | H&E | Gastric tumour segmentation | Patch-based FCN + iterative learning approach | China Big Data and AI challenge (1,900 images) |
| Ilse et al. (2018) (✓) | Multi-Cancers | H&E | Cancer image classification | MIL pooling based on gated-attention mechanism | CRCHisto (100 images) |
| Shujun Wang et al. (2019) | Stomach | H&E | Gastric cancer detection | Two-stage CNN framework for localization and classification | Private set (608 images) |
| Wang et al. (2019) | Lung | H&E | Lung cancer image classification | Patch based FCN + context-aware block selection and feature aggregation strategy | Private (939 WSIs), TCGA (500 WSIs) |
| Campanella et al. (2019) (✓) | Multi-Cancers | H&E | Multiple cancer diagnosis in WSIs | CNN (ResNet) + RNNs | Prostate (24,859 slides), skin (9,962 slides), breast cancer metastasis (9,894 slides) |
| Dov et al. (2019) | Thyroid | — | Thyroid malignancy prediction | CNN + ordinal regression for prediction of thyroid malignancy score | Private set (cytopathology 908 WSIs) |
| Xu et al. (2019) (✓) | Multi-Cancers | H&E | Segmentation of breast cancer metastasis and colon glands | FCN trained on instance-level labels, which are obtained from image-level annotations | Camelyon16 (400 WSIs), Colorectal adenoma private dataset (177 WSIs) |
| Huang and Chung (2019) | Breast | H&E | Localization of cancerous evidence in histopathology images | CNN + multi-branch attention modules and deep supervision mechanism | PCam (327,680 patches extracted from Camelyon16) and Camelyon16 (400 WSIs) |
| Other approaches | |||||
| Campanella et al. (2018) | Prostate | H&E | Prostate cancer detection | CNN trained under MIL formulation with top-1 ranked instance aggregation approach | Prostate biopsies (12,160 slides) |
| Akbar and Martel (2018) (✓) | Breast | H&E | Detection of breast cancer metastasis | Clustering (VAE + K-means) based MIL framework | Camelyon16 (400 WSIs) |
| Tellez et al. (2019b) (✓) | Multi-Cancers | H&E | Compression ofgigapixel histopathology WSIs | Unsupervised feature encoding method (VAE, Bi-GAN, contrastive training) that maps high-resolution image patches to low-dimensional embedding vectors | Camelyon16 (400 WSIs), TUPAC16 (492 WSIs), Rectum (74 WSIs) |
| Qu et al. (2019) (✓) | Multi-Cancers | H&E | Nuclei segmentation | Modified UNet trained using coarse level-labels + dense CRF loss for model refinement | MoNuSeg (30 images), lung cancer private set (40 images) |
| Bokhorst et al. (2019) | Colon | H&E | Segmentation of tissue types in colorectal cancer | UNet with modified loss functions to circumvent sparse manual annotations | Colorectal cancer WSIs (private set - 70 images) |
| Li et al. (2019a) (✓) | Breast | H&E | Mitosis detection | FCN trained with concentric loss on weakly annotated centriod label | ICPR12 (50 images), ICPR14 (1,696 images), AMIDA13 (606 images), TUPAC16 (107 images) |