Hi , Computational pathology applications using artificial intelligence are becoming increasingly complex. From detecting and classifying cells and tissue to predicting biomarkers and patient outcomes.
Simpler tasks rely upon pathologists’ annotations of specific features in the tissue. But biomarkers and outcomes are more complex. Algorithms must decipher large whole slide images without
any prior knowledge of which regions of tissue or characteristics of its appearance are important.
Risk stratification can already be done using cancer staging, molecular features, or clinical variables. However, improving prognostic insights is an active area of research.
The challenge with time-to-event data is that some patients do not have an event observed. Patients might not have died by the end of the study or they may have been lost to follow up during the study. These observations are known as right-censored.
A standard regression model is not suitable because the right-censored cases would need to be dropped, and a binary classification model (e.g., alive or dead after N years) would not make use of the actual survival times.
The most common solution is a Cox Proportional Hazards model that quantifies how much each covariate influences the survival rate. Some of the earlier work applying survival models to pathology images used hand-crafted features as the covariates in this linear model.
Deep learning models make use of this same formulation by applying a set of non-linear operations to
produce a feature set for the linear Cox model, with the whole model optimized end-to-end. To apply this model to images, the deep network is typically a CNN, and the output is a risk score.
The risk score produced by the model can be used to rank patients according to their level of risk. But sometimes these models are also used to stratify patients into groups with similar risk.
Survival model used by Meier et al. [3]
Muhammad et al. found that adding a second loss function to measure the accuracy of stratifying patients into high and low risk groups improved their model, especially on an independent test set [1].
Shirazi et al. abandoned the Cox model entirely, instead using a set of binary classifiers that divided the time axis into 4 classes: 0-6 months, 6-12 months, 12-24 months, and >24 months [2]. Other research has taken this multi-task approach further, using finer divisions of the time axis and additional constraints.
Other alternatives to the Cox model have been explored by Meier et al. who compared with the Uno and logrank losses, finding them slightly superior to the Cox Proportional Hazards model [3].
I’m interested to see these alternative survival loss functions compared in future work, as there doesn’t seem to yet be a consensus on which produces the best result.
Most survival models for histology have used H&E images
(including [1] and [2]), but [3] studied three different immunohistochemical stains. Other modalities like fluorescence are less explored.
