Portfolio Back to overview

Medical-image machine learning

Image-based direct survival prediction

Investigating why pooled survival results weakened within individual cohorts.

A learned image representation feeds a discrete-time survival output, conditional hazards, survival probabilities and a risk score.
A discrete-time survival model converts image features into conditional hazards and a patient risk score.

A discrete-time survival model converts image features into conditional hazards and a patient risk score.

I used image representations to predict conditional event probabilities across time intervals, then derived survival probabilities and risk scores. Evaluation included the concordance index, hazard ratios and Kaplan–Meier risk-group comparisons.

Looking inside the pooled result

UNI and prov-GigaPath reached pooled test C-indices of 0.618 and 0.626. Performance weakened after stratifying by TCGA, MD Anderson and POG, and predicted risk differed substantially by cohort. The stronger pooled result did not establish reliable discrimination within each cohort. I examined the distribution of risk scores and associations with molecular biomarkers. These associations were weak and sensitive to cohort composition. The work illustrates why external evaluation needs to look at cohort effects as well as aggregate performance.

Four Kaplan–Meier plots comparing low-risk and high-risk groups, with shaded confidence intervals, hazard ratios and log-rank p-values. The lower two panels show greater separation between groups.
Kaplan–Meier curves comparing predicted low-risk and high-risk groups across four encoders, with hazard ratios and log-rank p-values.

Kaplan–Meier curves comparing predicted low-risk and high-risk groups across four encoders, with hazard ratios and log-rank p-values.

Boxplots of predicted risk scores for TCGA, MDA and POG, showing the lowest median in MDA and the highest in POG, with annotated pairwise comparisons.
Predicted risk scores varied by cohort, with a lower median in MDA and a higher median in POG.

Predicted risk scores varied by cohort, with a lower median in MDA and a higher median in POG.

  • Multiple-Instance Learning
  • Domain Shift
  • Model Evaluation
Next projectCancer subtype reference stability