Virtual Library

 x 
My Library Virtual Library FAQ

Start Your Search

G. Smith



Author of

  • +

    MA 14 - Diagnostic Radiology, Staging and Screening for Lung Cancer I (ID 672)

    • Event: WCLC 2017
    • Type: Mini Oral
    • Track: Radiology/Staging/Screening
    • Presentations: 1
    • +

      MA 14.13 - Nodule Size Isn't Everything: Imaging Features Other Than Size Contribute to AI Based Risk Stratification of Solid Nodules (ID 8177)

      17:05 - 17:10  |  Author(s): G. Smith

      • Abstract
      • Presentation
      • Slides

      Background:
      Previously proposed risk models for the malignancy of Indeterminate Pulmonary Nodules (IPNs) detected on Computed Tomography (CT) typically incorporate a mixture of clinical factors, such as age and smoking history, and radiological factors such as nodule size and location. Of the latter, size is considered one of the most significant. Artificial Intelligence based risk stratification software has been previously proposed that uses Texture Analysis with Machine Learning to predict IPN malignancy and has been shown to achieve high classification performance. While it is assumed that such techniques can capture image texture patterns that separate benignity from malignancy, such methods also intrinsically measure nodule size. The contribution of texture to classifier performance beyond size has not been studied and we seek to quantify this. We show, for the first time, the relative contributions of texture and size on the performance of Artificial Intelligence risk stratification of solid nodules.

      Method:
      Two datasets were created from the US National Lung Screening Trial (NLST). The first (A), comprising 640 solid nodules, was built to remove size as a discriminatory factor between benign and malignant; all malignant solid nodules between 4 and 20 mm diameter were selected, and for each, a benign solid nodule was selected that most closely matched it in diameter. Any malignant nodule for which an equivalently sized benign could not be found within 0.8 mm was rejected. Sizes were measured using automated volumetric segmentation. The second dataset (B), also comprising 640 subjects, included all malignant nodules in A but benign nodules were randomly selected following the empirical size distribution of the whole NLST dataset. Therefore, nodule size cannot be a discriminative factor in A but would be in B. Two nodule stratification algorithms were developed using Texture Analysis combined with Machine Learning (Support Vector Regression) integrating 20 variables including 3D Haralick, Gabor and Shape features, from A and B respectively using five-fold cross validation and the performance compared measuring Area-Under-the-Curve (AUC).

      Result:
      The average AUC for the algorithm trained on dataset A was 0.70 whereas using size alone on the same dataset gave an AUC of 0.50. The AUC was 0.91 for the algorithm trained on B.

      Conclusion:
      On this data, Texture Analysis with Machine Learning contributes 0.20 AUC points to classfication performance. Artificial Intelligence based risk classification can identify radiological features that are predictive of solid nodule malignancy that are independent of nodule size.

      Only 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, select "Add to Cart" and proceed to checkout. If you would like to become a member of IASLC, please click here.

      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.