Virtual Library

Start Your Search

K. Desai



Author of

  • +

    P1.06 - Poster Session/ Screening and Early Detection (ID 218)

    • Event: WCLC 2015
    • Type: Poster
    • Track: Screening and Early Detection
    • Presentations: 1
    • +

      P1.06-023 - Addition of Low Dose Computed Tomography Image-Features Improves Diagnostic Accuracy for Indeterminate Pulmonary Nodules (ID 1019)

      09:30 - 09:30  |  Author(s): K. Desai

      • Abstract
      • Slides

      Background:
      Lung cancer is the leading cause of cancer related deaths world-wide. While low dose computed tomography (LDCT) screening of the high risk patient population was recently shown to decrease deaths from lung cancer by 20%, LDCT also resulted in 18% over-diagnosis [c.f. Patz-E.-F.-JAMA-2003] with a positive predictive value of only 52.9% when a suspicious LDCT finding led to a biopsy [c.f. Church-T.-NEJM-368-2003]. We tested whether combining novel image-features (IF) with routinely collected baseline-features (BF) can improve the accuracy of diagnosing suspicious findings on baseline LDCT.

      Methods:
      This exploratory case-control study included N=123 (66-cancer, 57-no-cancer) high risk subjects with at least one suspicious finding (nodule >= 8mm [c.f. Lung-RADS-ACR-2014]) on baseline LDCT screening at Vanderbilt University on a VCT Discovery (GE-Healthcare, UK) or a Brilliance iCT 128 SP (Philips, Amsterdam) system. The cohort was randomly divided into a separate training-set (N=55, 32-cancer, 23-no-cancer) and a test-set (N=68, 34-cancer, 34-no-cancer). All model training and leave-one-out cross-validation were strictly restricted to the training-set. Performance was evaluated on the unseen test-set. Definitive lung cancer or no-cancer diagnosis, smoking history and at least 6 baseline-features (BF6) viz. age, family-history, pack-smoking-years, body-mass-index, nodule-location, nodule-size were recorded for all subjects. Baseline lung cancer predictions were generated by (a) using the Gould-model [c.f. Gould,M.-Chest-2007] and (b) fitting an Elastic-Net Regularized Generalized Linear Model (GLMnet [c.f. Zou-H.-Journ-Royal-Stats-Soc-B-2005]) to BF6. The final baseline model (“GLMnet:BF”) effectively utilized 4 baseline-features with the coefficients for age and body-mass-index shrunk to zero. New LDCT specific information was extracted by computing 589 intensity, shape, surface and texture features (IF589) [c.f. Aerts-H.-Nat-Comm-S2014, Way-T.-Med-Phys-2009] from a 3D volume-of-interest (VOI) encompassing a rough Graph-cuts [c.f. Li-K.-IEEE-PAMI-2006] segmentation for each suspicious nodule. A GLMnet was fit to all 595 features (BF6 and IF589) yielding a final enhanced model (“GLMnet:BF+IF”), which contained 12 features after GLMnet shrinkage : 10 IF related to VOI energy, nodule shape and surface statistics and image intensity variability and 2 BF (family-history, nodule-location).

      Results:
      Baseline AUC increase by 7.4% from 0.81 (Gould-model) and 0.80 (GLMnet:BF) to 0.87 (GLMnet:BF+IF). At 88% sensitivity, false positive rate reduced by 60% from 56% (Gould-model) and 44% (GLMnet:BF) to 18% (GLMnet:BF+IF); accuracy improved from 65% (Gould-model) and 71% (GLMnet:BF) to 84% (GLMnet:BF+IF). Fig.1 below shows more details: Figure 1



      Conclusion:
      This initial exploratory analysis showed that image-features extracted from suspicious LDCT findings may help reduce the number of unnecessary biopsies. Additional validation studies are warranted to determine the value of this structural imaging-based approach.

      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.