Author of

• 25953

  +

  P2.16 - Treatment of Early Stage/Localized Disease (Not CME Accredited Session) (ID 965)

  • Event: WCLC 2018
  • Type: Poster Viewing in the Exhibit Hall
  • Track:
  • Presentations: 1
  • Moderators:
  • Coordinates: 9/25/2018, 16:45 - 18:00, Exhibit Hall

  • 27257

    +

    P2.16-02 - Predicting Pathological Noninvasiveness in T1 Non-Small Cell Lung Cancer on Chest CT Scan Using Deep Learning Algorithm (ID 13752)

    16:45 - 18:00  |  Author(s): Sae Jung Na
    • Abstract
    • Slides

    Background
    

    Because of the risk of recurrence, lobectomy is indicated even for small lung cancers. If we could accurately predict and classify noninvasive small lung cancers on CT images using deep learning algorithm prior to surgery, a more refined selection of candidates who would benefit from limited resection without increasing the risk of recurrence would be possible. The purpose of this study was to evaluate the ability of an artificial intelligence (deep learning algorithm) to predict pathological noninvasiveness of T1 non-small cell lung cancer (NSCLC) using computed tomography (CT) images.

    a9ded1e5ce5d75814730bb4caaf49419 Method
    

    468 preoperative CT images of NSCLC smaller than 3 cm, resected at our institution from 2008 to 2015 were used to train and validate the deep learning algorithm. Noninvasiveness was defined as the absence of nodal involvement, vascular invasion, and lymphatic invasion, and NSCLCs were classified as either noninvasive or invasive according to the pathological reports. A deep 3D convolutional neural network (CNN) was trained using 5-fold cross-validation method. To normalize input data, we rescaled CT image to let one voxel size be represented as (1mm,1mm,1mm). Input cube size was (32mm, 32mm,32mm). Horizontal flip and vertical flip augmentation were applied, and max-pooling and dropout layer were used to avoid overfitting. Receiver operating characteristic (ROC) curves and areas under the curve (AUCs), accuracy, and sensitivity/specificity were used to assess the performance of 3D CNN. We also added tumor size as an external feature to the 3D CNN model and evaluated the performance of the combined model using Xgboost classifier.

    4c3880bb027f159e801041b1021e88e8 Result
    

    157 out of 486 samples (33.5%) were invasive. A subsample group composed of 10% of the data (32 noninvasive and 16 invasive samples) was retained for the validation set.

    The 3D CNN showed an AUC of 0.826, 77.08% accuracy, 68.7% sensitivity, and 81.2 % specificity. Adding tumor size to the 3D CNN model showed an AUC of 0.855, 81.25% accuracy, 87.5% sensitivity, and 78.1% specificity.

    8eea62084ca7e541d918e823422bd82e Conclusion
    

    Artificial intelligence can accurately predict pathological noninvasiveness in T1 size NSCLC on CT images.

    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.