Divide-and-Rule: Self-Supervised Learning for Survival Analysis in Colorectal Cancer

Christian Abbet, Inti Zlobec, Behzad Bozorgtabar, Jean-Philippe Thiran

TL;DR

Researchers have developed a "Divide-and-Rule" (DnR) self-supervised learning framework that automatically discovers complex tissue patterns in colorectal cancer (CRC) histology slides. By learning how different tissue regions interact without needing manual expert labels, the model significantly outperforms standard pathology reports and state-of-the-art deep clustering methods in predicting patient survival.

What Problem Does This Paper Address?

In colorectal cancer, a patient’s prognosis is typically determined by a pathologist looking for about ten specific features (like tumor depth or lymph node metastasis). However, a single Whole Slide Image (WSI) contains a "snapshot" of millions of cells and complex microenvironments that the human eye might overlook.

The challenge is twofold:

  1. Label Scarcity: Manually labeling every tissue type in these massive images is prohibitively expensive for experts.
  2. Complexity: Existing automated methods often rely on simple features or noisy pre-trained networks that don't capture the spatial "dialogue" between different tissue types (e.g., how the tumor interacts with the surrounding stroma).

Methods

The authors propose a multi-stage pipeline called Divide-and-Rule (DnR):

  • RoI Detection: They first filter the slides to focus only on cancerous regions (epithelium, stroma, and lymphocytes) using a ResNet-18 classifier trained on a public CRC dataset.
  • Self-Supervised Representation: To learn features without labels, they use a "transfer colorization" proxy task. The model learns to reconstruct an image by separating it into its Hematoxylin and Eosin (HE) stain components, forcing the encoder to understand structural morphology.
  • The DnR Algorithm:
    • Divide: An instance loss ensures that overlapping image patches have similar embeddings, while distant patches are pushed apart.
  • Rule: A clustering loss groups similar patches together. It uses an entropy-based threshold that gradually expands, starting with "easy" similar samples and moving to "harder" ones to refine the cluster boundaries.
  • Survival Modeling: After clustering the tissue patterns, they represent each patient by the distribution of these clusters and—crucially—the transition probabilities between them (how often tissue A sits next to tissue B). These features are then fed into a Cox proportional hazards model.

Datasets

  • Primary Dataset: A retrospective collection of 660 in-house unlabeled WSIs from 374 unique CRC patients, including survival data and treatment info.
  • Pre-training Dataset: The "Kather 100K" dataset (100,000 histological images of CRC) was used specifically to train the initial Region of Interest (RoI) detector.

Key Insights and Ideas

  • Spatial Context is King: The model doesn't just look at what is in the slide, but where it is. By modeling the interaction between clusters (the "transition" features), the researchers captured prognostic information that single-feature analysis misses.
  • Color as a Proxy for Structure: By forcing the network to perform colorization (mapping RGB to HE stains), the model learns to identify nuclei and extracellular material without being told what they are.
  • Overfitting Mitigation: Survival analysis often suffers from having "small" datasets (hundreds of patients) relative to "large" feature sets. The authors used forward variable selection and leave-one-out cross-validation (LOOCV) to ensure their discovered patterns were statistically robust.

Key Results

  • Superior Prognosis: The DnR approach achieved a Concordance Index (C-Index) of 0.6943, significantly outperforming standard histopathological features (0.6076) and other deep clustering baselines like DEC (0.6426).
  • Discovery of "Subtle" Interactions: The model identified 13 specific features linked to survival. For example, it found that the specific interaction between blood vessels and tumor stroma is a strong indicator of lower survival outcomes.
  • Meaningful Clusters: When visualized, the unsupervised clusters corresponded to distinct biological entities: inflammatory tissues, muscle, large vessels, and even "edge" regions of the slide.

Limitations and Future Work

  • Dataset Size: While 374 patients is substantial for this type of specialized study, the authors acknowledge that larger, multi-center cohorts would be needed to further validate the clinical utility of these specific discovered patterns.
  • Black Box Nature: While the clusters can be visualized, the exact biological "reasoning" behind why certain tissue transitions predict death remains an area for deeper collaboration between AI and pathologists.
  • Generalization: The current model is tuned for CRC; future work could explore if the DnR framework generalizes to other cancer types like breast or lung cancer.

Summary updated

Abstract

With the long-term rapid increase in incidences of colorectal cancer (CRC), there is an urgent clinical need to improve risk stratification. The conventional pathology report is usually limited to only a few histopathological features. However, most of the tumor microenvironments used to describe patterns of aggressive tumor behavior are ignored. In this work, we aim to learn histopathological patterns within cancerous tissue regions that can be used to improve prognostic stratification for colorectal cancer. To do so, we propose a self-supervised learning method that jointly learns a representation of tissue regions as well as a metric of the clustering to obtain their underlying patterns. These histopathological patterns are then used to represent the interaction between complex tissues and predict clinical outcomes directly. We furthermore show that the proposed approach can benefit from linear predictors to avoid overfitting in patient outcomes predictions. To this end, we introduce a new well-characterized clinicopathological dataset, including a retrospective collective of 374 patients, with their survival time and treatment information. Histomorphological clusters obtained by our method are evaluated by training survival models. The experimental results demonstrate statistically significant patient stratification, and our approach outperformed the state-of-the-art deep clustering methods.
BibTeX
@inproceedings{abbet2020divide,
  author = {Christian Abbet and Inti Zlobec and Behzad Bozorgtabar and Jean-Philippe Thiran},
  title = {{Divide-and-Rule: Self-Supervised Learning for Survival Analysis in Colorectal Cancer}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2020},
  url = {https://openreview.net/forum?id=BNnH7kUw0F},
}
Figure