Image-Based Molecular Phenotyping of Pancreatic Ductal Adenocarcinoma

Georgios Kaissis, Sebastian Ziegelmayer, Fabian Lohöfer, F Harder, Friederike Jungmann, Daniel Sasse, Alexander Muckenhuber, Hsi-Yu Yen, Katja Steiger, Jens T. Siveke, Helmut Friess, Roland M. Schmid, Wilko Weichert, Marcus R. Makowski, Rickmer BrarenJournal of Clinical Medicine 2020 · DOI

TL;DR

This paper presents a machine learning model that uses radiomic features from CT scans to predict molecular subtypes of pancreatic ductal adenocarcinoma (PDAC). This non-invasive approach can stratify patients with previously unclassifiable tumors into groups with significantly different survival outcomes, paving the way for more personalized treatment strategies.

What Problem Does This Paper Address?

Pancreatic ductal adenocarcinoma (PDAC) is a deadly cancer, and recent research has identified distinct molecular subtypes with different prognoses and treatment responses. However, translating these discoveries into clinical practice is challenging due to tumor heterogeneity and the limitations of traditional biopsy methods, which only sample a small part of the tumor. This creates a gap between understanding molecular subtypes and improving patient outcomes. The authors aim to bridge this gap by developing a non-invasive method to predict these molecular subtypes using readily available CT imaging data, enabling better patient stratification and potentially guiding treatment decisions.

Methods

The core of this study involves a two-step process: radiomic feature extraction and machine learning classification.

  1. Radiomic Feature Extraction: The researchers used the Pyradiomics software to extract a large number (1474) of quantitative features from segmented CT scans of PDAC tumors. These features capture various aspects of the tumor's texture, shape, and intensity. After initial filtering for missing values, zero variance, and inter-observer variability (using an intra-class correlation coefficient < 0.9), 161 features were retained.
  2. Machine Learning Classification: A Random Forest algorithm was employed to build a model that predicts the molecular subtype (quasi-mesenchymal (QM) vs. non-quasi-mesenchymal (non-QM)) based on the radiomic features. To handle class imbalance, per-sample weighting was used. The model was trained and cross-validated using a five-fold approach. The performance was evaluated using sensitivity, specificity, and ROC-AUC.

Datasets

The study utilized a retrospective cohort of 207 patients with histologically confirmed pancreatic ductal adenocarcinoma.

  • Cohort A (Training/Cross-validation): 181 patients with unequivocally classified QM or non-QM molecular subtypes based on immunohistochemical staining (KRT81 and HNF1a).
  • Cohort B (Testing): 26 patients with histopathologically unclassifiable "double positive" (KRT81+/HNF1a+) tumors. This cohort was used to test the generalizability of the trained model.

The study did not introduce new benchmark datasets but rather used existing patient data for training and validation.

Key Insights and Ideas

  • Radiomics as a Non-Invasive Molecular Proxy: The study demonstrates that quantitative imaging features (radiomics) extracted from CT scans can serve as a powerful, non-invasive proxy for complex molecular subtypes of PDAC. This bypasses the need for invasive biopsies, which are often unreliable due to tumor heterogeneity.
  • Bridging the Translational Gap: By linking imaging phenotypes to molecular subtypes and then to clinical outcomes, the research directly addresses the translational gap between basic science discoveries and tangible patient benefits.
  • Unclassifiable Tumors Hold Prognostic Value: The most striking insight is the ability of the radiomics model to stratify patients with histopathologically "unclassifiable" tumors into two distinct groups with significantly different survival rates. This suggests that these tumors are not a single entity but likely represent distinct biological processes that can be identified through imaging.
  • Heterogeneity as a Key Feature: The study found that features related to image homogeneity/heterogeneity (like Entropy, Variance, Uniformity) were among the most important predictors. This highlights that the spatial variations within a tumor, visible on CT, are biologically significant.

Key Results

  • Classification Performance: The Random Forest model achieved high performance in classifying QM vs. non-QM tumors in the training cohort, with a sensitivity of 0.84 ± 0.05, specificity of 0.92 ± 0.01, and ROC-AUC of 0.93 ± 0.01.
  • Survival Prediction (Cohort A): For the initial cohort, predicted QM tumors had a median overall survival of 16.1 months, compared to 20.9 months for predicted non-QM tumors (log-rank test p = 0.02, HR 1.59).
  • Survival Prediction (Cohort B - Unclassifiable Tumors): When applied to the histopathologically unclassifiable tumors, the model identified two groups with dramatically different survival outcomes: 8.9 months for predicted QM and 39.8 months for predicted non-QM (log-rank test p < 0.001, HR 4.33). This is a highly significant finding.

Limitations and Future Work

The authors acknowledge several limitations:

  • Limited Dataset Size and Retrospective Nature: The study used a retrospective, single-institution dataset, which may limit the generalizability of the findings.
  • Potential for Overfitting: The retrospective nature and cross-validation approach could lead to overly optimistic classification metrics.
  • Lack of Multi-institutional Validation: Prospective validation across multiple institutions is needed to confirm the robustness of the algorithm.
  • Definition of Subtypes: The exact definition and robustness of PDAC subtypes, and how well histopathology represents transcriptome-based phenotypes, remain areas for further investigation.

Future work should focus on:

  • Larger, Multi-institutional Studies: Expanding the dataset and validating the model in diverse patient populations.
  • Prospective Validation: Conducting prospective studies to confirm the clinical utility and impact on patient outcomes.
  • Investigating Heterogeneity Biomarkers: Further exploring specific heterogeneity imaging biomarkers for QM tumors.
  • Integration with Other Modalities: Potentially combining CT radiomics with other imaging modalities or clinical data for even more precise stratification.

Summary updated

Abstract

To bridge the translational gap between recent discoveries of distinct molecular phenotypes of pancreatic cancer and tangible improvements in patient outcome, there is an urgent need to develop strategies and tools informing and improving the clinical decision process. Radiomics and machine learning approaches can offer non-invasive whole tumor analytics for clinical imaging data-based classification. The retrospective study assessed baseline computed tomography (CT) from 207 patients with proven pancreatic ductal adenocarcinoma (PDAC). Following expert level manual annotation, Pyradiomics was used for the extraction of 1474 radiomic features. The molecular tumor subtype was defined by immunohistochemical staining for KRT81 and HNF1a as quasi-mesenchymal (QM) vs. non-quasi-mesenchymal (non-QM). A Random Forest machine learning algorithm was developed to predict the molecular subtype from the radiomic features. The algorithm was then applied to an independent cohort of histopathologically unclassifiable tumors with distinct clinical outcomes. The classification algorithm achieved a sensitivity, specificity and ROC-AUC (area under the receiver operating characteristic curve) of 0.84 ± 0.05, 0.92 ± 0.01 and 0.93 ± 0.01, respectively. The median overall survival for predicted QM and non-QM tumors was 16.1 and 20.9 months, respectively, log-rank-test <i>p</i> = 0.02, harzard ratio (HR) 1.59. The application of the algorithm to histopathologically unclassifiable tumors revealed two groups with significantly different survival (8.9 and 39.8 months, log-rank-test <i>p</i> < 0.001, HR 4.33). The machine learning-based analysis of preoperative (CT) imaging allows the prediction of molecular PDAC subtypes highly relevant for patient survival, allowing advanced pre-operative patient stratification for precision medicine applications.
BibTeX
@article{kaissis2020image,
  author = {Georgios Kaissis and Sebastian Ziegelmayer and Fabian Lohöfer and F Harder and Friederike Jungmann and Daniel Sasse and Alexander Muckenhuber and Hsi-Yu Yen and Katja Steiger and Jens T. Siveke and Helmut Friess and Roland M. Schmid and Wilko Weichert and Marcus R. Makowski and Rickmer Braren},
  title = {{Image-Based Molecular Phenotyping of Pancreatic Ductal Adenocarcinoma}},
  journal = {Journal of Clinical Medicine},
  year = {2020},
  doi = {10.3390/jcm9030724},
}
Figure