Fig 1.
Recurrence-free survival for the colorectal carcinoma patients—(a) internal and (b) external sites.
P-values are computed as a log-rank test using the internal site Ontario as a reference (ref).
Table 1.
Description of the dataset—Shows the characteristics of both internal and external datasets.
Fig 2.
Quantitative feature extraction pipeline on 2 sample images, which segments the image in a stepwise manner.
First, the image is segmented into carcinoma (green), stroma (light blue), mucin (dark blue), TB/PDC (red), necrosis (brown), smooth muscle (purple, and fat (yellow). Next, the stroma is segmented into immature (teal), mature (green), and inflammatory (gray). The carcinoma is segmented into low-grade (purple), high-grade (orange), and signet ring cells (light green). Finally, TILs are recognized as objects (blue dots) within the tumor. After this segmentation, 15 features are calculated from each image as shown. Abbreviations: B, tumor bed; ST, stromal region.
Fig 3.
Causal modeling framework: (left) latent shift assumption and causal relations; (right) proposed model diagram demonstrates 3 main components of the model.
A) learns a latent representation capturing task relevant information alongside proxy information. B) Attempts to infer the latent variables directly from the input features. C) A risk estimation model is trained, predictions are modified using the latent estimates.
Fig 4.
Pearson correlation coefficient between Stage (C), Center (W), QuantCRC features (X) and target recurrence (Y)—(a) represents stage and center as separate variables, (b) represents stage and center as a combined variables, and (c) Correlation with recurrence.
Table 2.
Performance of the models on the internal and external test sets in-terms of C-index.
95% confidence interval is calculated using auto-bootstraping. Bold front represents optimal performance.
Fig 5.
Comparative area under the receiver operating characteristics curve (AUROC) calculated for every time step on the internal and external dataset—(left) internal and (right) external.
Blue: Baseline Cox model, Orange: DeepSurv, and Green: Proposed Causal Survival model.
Table 3.
Configuration of ablation study.
Removed components of the causal survival model is marked as X. Performance measured as C-index with 95% confidence interval.