Table 6:
Overview of survival models for disease prognosis. Note: (✓) indicates the code is publicly available and the link is provided in their respective paper.
| Reference | Cancer types | Application | Method | Dataset |
|---|---|---|---|---|
| Zhu et al. (2017b) | Multi-Cancers | Loss function based on survival time | Raw pixel values of downsampled patches used as feature vectors; 10 clusters identified using K-means clustering. Deep survival models are trained for each cluster separately. Significant clusters are identified and corresponding scores are fed into final WSI classifier | TCIA-NLST, TCGA-LUSC, TCGA-GBM |
| Bychkov et al. (2018) | Colorectal | 5 year disease specific survival | Extracted features using pre-trained VGG-16. Used RNN to generate WSI prediction from tiles | Private set - TMAs from 420 patients |
| Couture et al. (2018) (✓) | Breast | Prediction of tumour grade, ER status, PAM50 intrinsic subtype, histologic subtype and risk of recurrence score | Pre-trained VGG-16 model. Aggregate features over 800 × 800 regions to predict class for each patch, then frequency distribution of classes input to SVM to combine regions to predict TMA class | TMA cores (Private-1203 cases) |
| Mobadersany et al. (2018) (✓) | Brain | Time to event modelling | CNN integrated with a Cox proportional hazards model to predict patient outcomes using histology and genomic biomarkers. Calculate median risk for each ROI, then average 2 highest risk regions | TCGA-LGG, TCGA-GBM (1,061 WSIs) |
| Courtiol et al. (2019) | Mesothelioma | Loss function based on survival time | Pre-trained ResNet50 extracts features from 10000 tiles. 1-D convolutional layer generates score for each tile. 10 highest and lowest scores fed into MLP classifier for WSI prediction | MESOPATH/MESOBANK (private set-2,981 WSIs), TCGA validation set (56 WSIs) |
| Geessink et al. (2019) | Colorectal | Dichotomized tumour/stromal ratios | CNN based patch classifier trained to identify tissue components. Calculate tumour-stroma ratio for manually defined hot-spots | Private set-129 WSIs |
| Kather et al. (2019) (✓) | Colorectal | Dichotomized stromal score | VGG-19 based patch classifier trained to identify tissue component. Calculate HR for each tissue component using mean activation. Combine components with HR > 1 to give a “deep stromal score” | NCT-CRC-HE-100k; TCGA-READ, TCGA-COAD |
| Muhammad et al. (2019) | Liver ICC | HRs of clusters compared | Unsupervised method to cluster tiles using autoencoder. WSI assigned to cluster corresponding to majority of tiles | Private set - 246 ICC H&E WSIs |
| Nagpal et al. (2019) | Prostate | Gleason scoring | Trained Inception-V3 network to predict Gleason score on labeled patches. Then calculate % patches with each grade on the WSI and use result as a low dimensional feature vector input to k-NN classifier | TCGA-PRAD and private dataset |
| Qaiser et al. (2019a) | Lymphoma | Generate 4 DPC categories | Multi-task CNN model for simultaneous cell detection and classification, followed by digital proximity signature (DPS) estimation | Private set-32 IHC WSIs |
| Tang et al. (2019) | Multi-Cancers | Dichototomized survival time (<= 1 year and > 1 year) | A capsule network is trained using a loss function that combines a reconstruction loss, margin loss ans Cox loss. The mean of all patch-level survival predictions is calculated to achieve a final patient-level survival prediction. | TCGA-GBM and TCGA-LUSC |
| Veta et al. (2019) | Breast | Predict mitotic score & PAM50 proliferation score | Multiple methods from challenge teams | TUPAC 2016 |
| Yamamoto et al. (2019) | Prostate | Predict accuracy of prostate cancer recurrence | Deep autoencoders trained at different magnifications and weighted non-hierarchical clustering, followed by SVM classifier to predict the short-term biochemical recurrence of prostate cancer | Private set - 15,464 WSI’s |