[Paper Review] Towards Unsupervised Cancer Subtyping: Predicting Prognosis Using A Histologic Visual Dictionary

Muhammad Hassan, Carlie Sigel+10 more|arXiv (Cornell University)|Mar 12, 2019

Cholangiocarcinoma and Gallbladder Cancer Studies25 references10 citations

TL;DR

This paper proposes an unsupervised deep convolutional autoencoder model that learns a histologic visual dictionary from whole-slide images of intrahepatic cholangiocarcinoma (ICC), enabling clustering of tumor morphologies without prior labeling. The method identifies three prognostic clusters linked to recurrence-free survival, with stromal fibrosis emerging as a key morphologic predictor, demonstrating computational pathology's potential for rare cancer subtyping where traditional grading is absent.

ABSTRACT

Unlike common cancers, such as those of the prostate and breast, tumor grading in rare cancers is difficult and largely undefined because of small sample sizes, the sheer volume of time needed to undertake on such a task, and the inherent difficulty of extracting human-observed patterns. One of the most challenging examples is intrahepatic cholangiocarcinoma (ICC), a primary liver cancer arising from the biliary system, for which there is well-recognized tumor heterogeneity and no grading paradigm or prognostic biomarkers. In this paper, we propose a new unsupervised deep convolutional autoencoder-based clustering model that groups together cellular and structural morphologies of tumor in 246 ICC digitized whole slides, based on visual similarity. From this visual dictionary of histologic patterns, we use the clusters as covariates to train Cox-proportional hazard survival models. In univariate analysis, three clusters were significantly associated with recurrence-free survival. Combinations of these clusters were significant in multivariate analysis. In a multivariate analysis of all clusters, five showed significance to recurrence-free survival, however the overall model was not measured to be significant. Finally, a pathologist assigned clinical terminology to the significant clusters in the visual dictionary and found evidence supporting the hypothesis that collagen-enriched fibrosis plays a role in disease severity. These results offer insight into the future of cancer subtyping and show that computational pathology can contribute to disease prognostication, especially in rare cancers.

Motivation & Objective

• To address the lack of standardized histopathology-based subtyping and prognostic biomarkers in rare cancers like intrahepatic cholangiocarcinoma (ICC).
• To develop an unsupervised deep learning framework that identifies biologically relevant morphologic patterns from digitized whole-slide images without manual annotation.
• To evaluate whether clustering-based morphologic subtypes derived from a visual dictionary can predict recurrence-free survival in ICC.
• To translate computational clusters into clinically interpretable histopathological terms to validate biological relevance.

Proposed method

• Trained a deep convolutional autoencoder on 246 whole-slide images of ICC to learn compact, hierarchical feature representations of histologic patterns.
• Applied k-means clustering to the learned latent representations to generate a visual dictionary of 24 clusters representing distinct morphologic phenotypes.
• Used the cluster assignments as covariates in univariate and multivariate Cox proportional hazards models to assess association with recurrence-free survival.
• Performed pathologist-led interpretation of significant clusters to assign clinical terminology and assess morphologic features.
• Evaluated model performance using likelihood ratio, Wald, and log-rank tests for survival significance.
• Quantified major and minor histologic features (e.g., tumor content, stroma, collagen) across clusters using semi-quantitative tile-level analysis.

Experimental results

Research questions

• RQ1 Can an unsupervised deep learning model identify biologically meaningful morphologic subtypes in ICC without prior labeling or clinical annotations?
• RQ2 Which histologic patterns, as defined by the visual dictionary, are significantly associated with recurrence-free survival in ICC?
• RQ3 Can the morphologic features of significant clusters be interpreted using standard histopathological terminology and linked to known biological processes?
• RQ4 Does tumor stroma, particularly collagen-rich fibrosis, exhibit morphologic heterogeneity that correlates with prognosis in ICC?
• RQ5 Can a computational visual dictionary serve as a foundation for developing prognostic models in rare cancers lacking established grading systems?

Key findings

• Three clusters (0, 11, 13) showed significant association with recurrence-free survival in univariate Cox regression, with hazard ratios of 0.618***, 0.515**, and 0.750*, respectively.
• In multivariate analysis, combinations of clusters 0, 11, and 13 remained significant, with the full model achieving a likelihood ratio test p-value of 11.37***.
• Cluster 0, characterized by >50% solid tumor with low nuclear:cytoplasmic ratio, was associated with reduced risk (HR = 0.618***), indicating a favorable prognosis.
• Cluster 11, defined by high stromal content and collagen deposition, showed a strong protective effect (HR = 0.515**), suggesting fibrotic stroma may correlate with better outcomes.
• Cluster 13, also enriched in stroma and collagen, was associated with reduced risk (HR = 0.750*), further implicating the tumor microenvironment in prognosis.
• Pathologist review confirmed that collagen-enriched fibrosis was a major/minor feature in two significant clusters, supporting its potential role in disease severity and prognosis.