Computer‐extracted features of nuclear morphology in hematoxylin and eosin images distinguish <scp>s</scp>tage <scp>II</scp> and <scp>IV</scp> colon tumors

Neeraj Kumar, Ruchika Verma, Chuheng Chen, Cheng Lu, Pingfu Fu, Joseph Willis, Anant MadabhushiThe Journal of Pathology 2022 · DOI

TL;DR

This paper introduces a computational pathology approach using deep learning to extract quantitative nuclear features from H&E stained colon cancer images. These features effectively distinguish between stage II (less aggressive) and stage IV (metastatic) colon tumors, and even predict patient survival, offering a more objective and reproducible method for risk stratification.

What Problem Does This Paper Address?

Accurately staging colorectal cancer (CRC) is crucial for determining the best treatment and predicting patient outcomes. While the TNM staging system is the current standard, it has limitations, particularly for stage II patients, where a significant percentage (up to 25%) can still develop distant metastases within 10 years. Current methods for identifying high-risk patients often rely on subjective interpretations of histological features or expensive molecular markers, leading to challenges in reproducibility and clinical adoption. There's a clear need for more objective, quantitative, and readily available tools to improve the prognostication of CRC patients, especially to identify those at risk of recurrence or metastasis.

Methods

The study employs a multi-stage computational pipeline:

  1. Tumor Segmentation: A fully convolutional neural network (FCN) with a VGG-18 architecture is trained to accurately identify and delineate tumor regions within whole slide images (WSIs). This step ensures that subsequent analysis is focused on cancerous tissue.
  2. Nuclei Segmentation: A Mask R-CNN model, leveraging a ResNet-50 backbone, is used to precisely segment individual nuclei within the identified tumor regions. This segmentation is crucial for extracting detailed morphological features from each cell.
  3. Feature Extraction: A comprehensive set of 26,641 quantitative morphometric features are extracted from each segmented nucleus. These features capture various aspects of nuclear morphology, including size (e.g., area, perimeter), shape (e.g., invariant moments, Fourier descriptors), texture (e.g., Haralick features), and orientation.
  4. Feature Selection and Classification: A Wilcoxon rank-sum test is used to identify the top five most discriminatory features between stage II and stage IV colon cancers. These selected features are then used to train a random forest classifier. This approach aims to balance predictive power with interpretability and reduce the risk of overfitting.
  5. Survival Analysis: For a subset of patients with survival data, Cox proportional hazards models and Kaplan-Meier curves are used to assess the prognostic value of the identified nuclear features and the classifier's output.
  6. Unsupervised Clustering: UMAP and hierarchical clustering are employed to explore the relationships between different tumor subtypes (stage II, stage IV with peritoneal spread, stage IV with hematogenous spread) based on the top nuclear features.

Datasets

  • Discovery Cohort: 100 cases of stage II and stage IV colon cancer from the University Hospitals Cleveland Medical Center (UHCMC). This cohort was split into a training set (100 cases) and a validation set (51 stage II, 79 stage IV).
  • Independent Validation Cohort: The Cancer Genome Atlas's Colon Adenocarcinoma (TCGA-COAD) dataset, comprising 143 stage II and 54 stage IV colon cancer cases. This dataset was used for external validation of the classification model and for survival analysis.
  • Additional Validation Set: 21 stage IV colon cancer cases with peritoneal metastases were added to the UHCMC validation set to specifically assess their similarity to stage II tumors.

Key Insights and Ideas

  • Morphological Signatures of Metastasis: The study demonstrates that subtle, quantitative differences in nuclear morphology can reliably distinguish between localized (stage II) and metastatic (stage IV) colon cancers. This suggests that the cellular changes associated with metastasis are reflected in the physical characteristics of the nucleus.
  • Interpretability Matters: By focusing on "handcrafted" features derived from established morphological principles rather than purely end-to-end deep learning, the authors achieve a more interpretable model. This is crucial for clinical adoption, as it allows pathologists to understand why a prediction is made.
  • Peritoneal vs. Hematogenous Spread: A particularly interesting finding is that stage IV colon cancers with peritoneal spread exhibit nuclear features that are more similar to stage II tumors than to stage IV tumors with hematogenous spread. This suggests distinct underlying biological processes or cellular adaptations for different metastatic pathways.
  • Prognostic Power Beyond Staging: The identified nuclear features not only distinguish between stages but also show a significant association with overall patient survival, highlighting their potential as independent prognostic markers.

Key Results

  • Classification Accuracy: The random forest classifier, using the top five discriminatory nuclear features, achieved an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.81 on the UHCMC validation set and 0.78 on the TCGA-COAD validation set.
  • Survival Prediction: In the TCGA-COAD cohort, the top five features were used in a Cox proportional hazards model to stratify patients, yielding a hazard ratio of 2.20 (95% CI 1.24-3.88) and a concordance index of 0.71. Kaplan-Meier analysis showed a statistically significant separation between low-risk and high-risk groups (log-rank P value of 0.0097).
  • Discriminatory Features: The most significant features identified were related to nuclear size (area, perimeter, major-axis length), nuclear orientation entropy, and local nuclear contrast variance.
  • Subtype Similarity: Unsupervised clustering revealed that stage IV colon cancers with peritoneal spread clustered closer to stage II colon cancers than to stage IV colon cancers with hematogenous spread.

Limitations and Future Work

  • Retrospective Study: The study is retrospective, and the analysis was performed on a limited number of patients.
  • Limited Clinical Variables: Additional clinical variables (e.g., patient age, race, tumor grade) were not integrated into the multivariate analysis.
  • Multi-institutional Validation: The authors acknowledge the need for validation on larger, multi-institutional datasets to confirm the generalizability of their findings.
  • Prospective Validation: Future work will focus on prospective validation and exploring the use of these quantitative features to guide adjuvant chemotherapy decisions for stage II patients.
  • Adjuvant Therapy Prediction: The authors aim to investigate whether these features can predict the need for adjuvant chemotherapy in stage II CC patients.

Summary updated

Abstract

We assessed the utility of quantitative features of colon cancer nuclei, extracted from digitized hematoxylin and eosin-stained whole slide images (WSIs), to distinguish between stage II and stage IV colon cancers. Our discovery cohort comprised 100 stage II and stage IV colon cancer cases sourced from the University Hospitals Cleveland Medical Center (UHCMC). We performed initial (independent) model validation on 51 (143) stage II and 79 (54) stage IV colon cancer cases from UHCMC (The Cancer Genome Atlas's Colon Adenocarcinoma, TCGA-COAD, cohort). Our approach comprised the following steps: (1) a fully convolutional deep neural network with VGG-18 architecture was trained to locate cancer on WSIs; (2) another deep-learning model based on Mask-RCNN with Resnet-50 architecture was used to segment all nuclei from within the identified cancer region; (3) a total of 26 641 quantitative morphometric features pertaining to nuclear shape, size, and texture were extracted from within and outside tumor nuclei; (4) a random forest classifier was trained to distinguish between stage II and stage IV colon cancers using the five most discriminatory features selected by the Wilcoxon rank-sum test. Our trained classifier using these top five features yielded an AUC of 0.81 and 0.78, respectively, on the held-out cases in the UHCMC and TCGA validation sets. For 197 TCGA-COAD cases, the Cox proportional hazards model yielded a hazard ratio of 2.20 (95% CI 1.24-3.88) with a concordance index of 0.71, using only the top five features for risk stratification of overall survival. The Kaplan-Meier estimate also showed statistically significant separation between the low-risk and high-risk patients, with a log-rank P value of 0.0097. Finally, unsupervised clustering of the top five features revealed that stage IV colon cancers with peritoneal spread were morphologically more similar to stage II colon cancers with no long-term metastases than to stage IV colon cancers with hematogenous spread. © 2022 The Authors. The Journal of Pathology published by John Wiley & Sons Ltd on behalf of The Pathological Society of Great Britain and Ireland.
BibTeX
@article{kumar2022computerextracted,
  author = {Neeraj Kumar and Ruchika Verma and Chuheng Chen and Cheng Lu and Pingfu Fu and Joseph Willis and Anant Madabhushi},
  title = {{Computer‐extracted features of nuclear morphology in hematoxylin and eosin images distinguish <scp>s</scp>tage <scp>II</scp> and <scp>IV</scp> colon tumors}},
  journal = {The Journal of Pathology},
  year = {2022},
  doi = {10.1002/path.5864},
}
Figure