Censor-Aware Semi-supervised Learning for Survival Time Prediction from Medical Images

Renato Hermoza, Gabriel Maicas, Jacinto C. Nascimento, Gustavo Carneiro

Introduction

Survival analysis, the study of the time until an event of interest occurs, is a fundamental component of clinical prognosis and treatment planning. In medical imaging, the goal is often to predict the survival time of a patient based on diagnostic images such as pathology slides or chest X-rays. However, this task is uniquely challenged by "right-censoring." A data point is right-censored when the event of interest (e.g., death) has not occurred by the end of the study period or when a patient drops out. In these cases, we only know that the patient survived at least until a certain time $T$, but their actual survival duration remains unknown.

Traditional approaches to this problem generally fall into two categories: Cox Proportional Hazards models, which excel at ranking patient risk but fail to predict specific time durations, and standard regression models, which often discard censored data entirely. This paper introduces a novel semi-supervised framework that bridges this gap by treating censored data as valuable training samples rather than missing values, enabling direct survival time prediction with state-of-the-art accuracy.

Methodology

The proposed framework utilizes a ResNet-18 backbone to extract features from medical images. To handle the continuous nature of time, the model discretizes survival duration into discrete "bins." The core innovation of the research lies in its semi-supervised training strategy, which leverages both uncensored and censored data through three primary mechanisms:

1. Pseudo-Labeling for Censored Data

For censored patients known to have survived $Y$ days, the model generates a "pseudo-label" $\gamma$. This label is calculated as the maximum of the known censoring time and the model's current prediction. By assigning a specific (though estimated) survival time to censored cases, the authors transform the survival problem into a semi-supervised regression task. This allows the model to learn from the "healthy" features of long-term survivors who would otherwise be excluded from regression training.

2. Censor-Aware Loss Functions

To ensure the model respects the physical constraints of censoring, the authors developed the Censor-Aware Mean Squared Error (CA-MSE). Unlike standard MSE, which penalizes any deviation from a target, CA-MSE is asymmetric. If a patient is censored at day 100, the model is only penalized if it predicts a survival time less than 100. Predictions greater than 100 incur zero penalty, as they are consistent with the known lower bound of the patient's survival.

Additionally, a Censor-Aware Rank Loss is employed. This ensures that the model maintains a logical ordering of patients, penalizing the network if it predicts a longer life for a patient who died earlier than a comparison subject.

3. Early-Learning Regularization (ELR)

Because pseudo-labels are generated by the model itself, they can be noisy—especially in the early stages of training. To prevent the model from "memorizing" its own incorrect guesses, the authors adapted Early-Learning Regularization (ELR). This technique stabilizes training by encouraging the model to favor long-term temporal trends in the data over the immediate (and potentially erroneous) pseudo-labels provided in early epochs.

Experimental Results

The framework was evaluated on two large-scale medical imaging benchmarks representing different clinical challenges:

  • TCGA-GM (Pathology): Comprising 1,505 pathology patches from 769 patients with brain gliomas. This dataset features a relatively balanced split between censored and uncensored cases. The proposed method achieved a Mean Absolute Error (MAE) of 286.95 days, significantly outperforming the previous state-of-the-art (CDOR) of 321.2 days. It also reached a Concordance Index (C-index) of 0.740.
  • NLST (Chest X-ray): This dataset from the National Lung Screening Trial includes 47,947 images. It represents a "needle in a haystack" scenario where over 95% of the data is censored. Despite the extreme scarcity of ground-truth death dates, the model achieved an MAE of 26.28 days and a C-index of 0.756, demonstrating remarkable robustness to high censoring rates.

Ablation studies confirmed that the inclusion of pseudo-labels for censored data improved the MAE by more than 20% compared to baseline models that relied solely on ranking losses.

Discussion and Conclusion

This research demonstrates that censored data contains critical information for survival time prediction. By treating censoring as a lower-bound constraint rather than a missing value, the authors' semi-supervised approach allows deep learning models to utilize 100% of available clinical datasets.

While the method shows great promise, the authors acknowledge a "cold start" limitation: the model requires a minimum threshold of uncensored records to begin generating reliable pseudo-labels. Future work may explore the application of this framework to other time-to-event domains, such as predictive maintenance or customer churn, where right-censoring is equally prevalent. Ultimately, this censor-aware approach provides a more accurate and data-efficient pathway for personalized prognostic modeling in healthcare.

Summary updated

BibTeX
@inproceedings{hermoza2022censor,
  author = {Renato Hermoza and Gabriel Maicas and Jacinto C. Nascimento and Gustavo Carneiro},
  title = {{Censor-Aware Semi-supervised Learning for Survival Time Prediction from Medical Images}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2022},
  doi = {10.1007/978-3-031-16449-1\_21},
}
Figure