Author of

• 15382

  +

  P3.04 - Poster Session/ Biology, Pathology, and Molecular Testing (ID 235)

  • Event: WCLC 2015
  • Type: Poster
  • Track: Biology, Pathology, and Molecular Testing
  • Presentations: 1
  • Moderators:
  • Coordinates: 9/09/2015, 09:30 - 17:00, Exhibit Hall (Hall B+C)

  • 15413

    +

    P3.04-031 - Combining CT Texture Analysis with Semantic Imaging Descriptions for the Radiogenomic Detection of EGFR and KRAS Mutations in NSCLC (ID 2965)

    09:30 - 09:30  |  Author(s): F. Stingo
    • Abstract

    Background:
    Existing literature suggests quantitative texture features derived from CT imaging can differentiate tumor genotypes and phenotypes. We combined CT texture analysis with semantic imaging descriptions provided by radiologists, and evaluated their ability to identify EGFR and KRAS mutation status in NSCLC.
    
    Methods:
    We retrospectively reviewed CT images from 628 patients from the GEMINI (Genomic Marker-Guided Therapy Initiative) cohort. Included were NSCLC patients whose biopsies included genetic testing for EGFR or KRAS mutations, and who underwent contrast-enhanced CT imaging within 90 days of biopsy. Excluded were patients who had undergone therapy or biopsy of their primary tumor before imaging, or whose tumors weren’t segmentable. All CT images were contrast-enhanced, with body kernel reconstruction, and slice thicknesses of 1.25-5mm. Tumor segmentation was done in 3DSlicer (Harvard University, Cambridge MA) using a semi-automatic segmentation algorithm. Image pre-processing and textural feature extraction was performed using IBEX (MDACC, Houston TX). Semantic descriptions of the tumors were recorded by a thoracic radiology fellow and a board-certified thoracic radiologist in consensus. For each patient a set of textural features was calculated, based on the GreyLevel Co-Occurrence Matrix, Run-Length Matrix, voxel intensity histogram, and geometric properties of the tumor. Feature selection was based on existing literature, prior research experience, and excluded those features previously found to be poorly reproducible in lung tissue. These were combined with semantic descriptions (e.g. presence or absence of features such as spiculations, air bronchograms, and pleural effusions), for a total of 51 textural and geometric features, and 11 semantic features. When available, the SUVmax for the tumor was also included. To detect correlations with genetic mutations, these features were combined to train a Random Forest machine learning algorithm. This algorithm output a prediction for the mutation status of each tumor, and the predictive accuracy was assessed based on 10-fold cross-validation.
    
    Results:
    Included were 121 patients, 113 tested for KRAS mutations (26 positive) and 118 tested for EGFR mutations (31 positive). Maximum tumor dimensions ranged from 1.2–15.5cm (mean 5.6cm). Individual semantic features found to correlate with mutation status included tumor cavitation, pleural effusion, presence of ground glass opacity, and the nature of tumor margins (all p-values <0.05). Used collectively in a Random Forest classifier, textural features alone showed a sensitivity and specificity for KRAS detection of 50% and 81% respectively, with 74% overall accuracy. This increased modestly to a sensitivity and specificity of 50% and 84% respectively when semantic features were added, with accuracy increasing to 77%. For EGFR detection, textural features had sensitivity and specificity of 48% and 77% respectively, giving 69% accuracy. Detection of EGFR did not improve with inclusion of semantic features.
    
    Conclusion:
    Texture analysis correctly identified EGFR and KRAS mutation status in most patients. Although some semantic features correlated with mutation status, when combined with textural features they provided little or no improvement in predictive accuracy. One possible explanation is that textural features may already be capturing the information contained in the semantic features. Our results suggest oncogenic drivers of NSCLC are associated with distinct imaging features that can be detected radiographically.