Author of

• 25931

  +

  P1.11 - Screening and Early Detection (Not CME Accredited Session) (ID 943)

  • Event: WCLC 2018
  • Type: Poster Viewing in the Exhibit Hall
  • Track:
  • Presentations: 1
  • Moderators:
  • Coordinates: 9/24/2018, 16:45 - 18:00, Exhibit Hall

  • 26556

    +

    P1.11-10 - Optimizing Radiomics Features by Minimizing Boundary Effects and Normalizing with Opposite Lung Tissue Characteristics (ID 14062)

    16:45 - 18:00  |  Presenting Author(s): Saeed Seyyedi
    • Abstract
    • Slides

    Background
    

    For wide adoption of LDCT screening it is thought that CAD will likely by necessary. We hypothesize that CAD features that minimizes perimeter effects and normalizes nodule CT features using the lung parenchyma from the opposite lung will improve the ability to determine nodule malignancy.

    a9ded1e5ce5d75814730bb4caaf49419 Method
    

    We have developed a CAD system that includes lung tissue segmentation, nodule detection and feature extraction from the segmented nodule, the segmented nodule minus the perimeter transition pixels (Core), and the opposite lung parenchymal tissue. (See Figure 1).

    We use the Mann-Whitney U test to compare teh discriminating ability of individual features and combinations of features) extracted from the nodule, nodule core and nodule normalized by mirror region features.

    

    4c3880bb027f159e801041b1021e88e8 Result
    

    In total, 34 early small suspect baseline nodules detected as part of the PanCan screening trial were used, these include 17 nodules proven to be cancer and 17 nodules that resolved on follow-up scans. The comparison of classification ability of features from nodules without edge vs. nodules with edge pixels reveals that the core features show better classification ability for 76 out of the 136 calculated features.

    Performing a leave-one-out LDA classifier cross-validation approach in using core features, gives an accuracy of 76% with only 1 feature through 3 features, and 82% with 4 features. However, repeating the same experiment for core plus edge features, shows accuracy of 67% with only 1 feature, 73% with 3 features, 79% with 4 features. Normalizing the core texture features by the texture features of the mirrored region in the opposite lung shows an improved classification ability for 52 out of the 89 texture features.

    8eea62084ca7e541d918e823422bd82e Conclusion
    

    In this study, the results suggest using the nodule core improves feature classification as does normalizing of the nodule by the mirrored region in opposite lung.

    6f8b794f3246b0c1e1780bb4d4d5dc53

    Only Active Members that have purchased this event or have registered via an access code will be able to view this content. To view this presentation, please login or select "Add to Cart" and proceed to checkout.