Virtual Library

Start Your Search

Ian Janzen

Author of

  • +

    OA06 - Refining Lung Cancer Screening (ID 131)

    • Event: WCLC 2019
    • Type: Oral Session
    • Track: Screening and Early Detection
    • Presentations: 1
    • Now Available
    • +

      OA06.01 - Comparison Between Radiomics-Based Machine Learning and Deep Learning Image Classification for Sub-Cm Lung Nodules (Now Available) (ID 2810)

      11:00 - 12:30  |  Author(s): Ian Janzen

      • Abstract
      • Presentation
      • Slides


      New clinical challenges have arisen from the recent recognition for an improved mortality of cancers via lung cancer screening using LDCT. A particular challenge for physicians and CADx systems is the classification and prediction of behavior for sub-cm lung nodules that are frequently present in screening CT scans. By predicting and classifying the behavior of these small nodules, we can identify potential cancerous nodules into the earlier stages of malignancy making them more easily treatable.


      We have evaluated a multitude of image processing techniques to assist in CADx systems for these small nodules such as Radiomic feature-based machine learning algorithms (linear discriminant analysis) as well as leveraging pretrained convolution neural networks such as VGG19 and InceptionV3 using deep learning/transfer learning techniques. The linear discriminate Radiomic analysis (LDA) classified a sample of CT imaged nodules (n=514) using quasi-volumetric nodule data (images of the nodules from CT slices above and below the central slice) into three discriminate categories: cancerous (clinically confirmed, n = 140) versus resolved (not present in follow up CT scans, n=107) versus stable (a negligible change in shape, texture, size in multi year follow up CT scans, n=267). Each nodule was segmented from the original CT scan using an inhouse lung CT image segmenation routine. This routine generated 2167 discrete CT nodule images upon which 133 Radiomic shape and texture features were calculated.


      The LDA Radiomic analysis correctly classified the individual nodual sections with an accuracy of 75.1% (jackknife - leave one out result) using only 18 features predefined traditional image analysis features (4 shape feature(s), 14 texture feature(s)) for cancer vs resolved + stable nodules. Requiring that more than or equal to 50% of sections from a nodule be classified as cancer for the nodule to be classified as cancer individual nodules could be correctly classified with an 80% accuracy.

      The leveraged pretrained networks (VGG19, and InceptionV3) trained using standard data augmentation and finetuning techniques, trained on this same quasi-volumetric image data as a binary classification task (malignant vs. benign nodules) achieved an average classification accuracy of 71% and 75% respectively through 10-crossfold validation.


      Machine learning using 18 Radiomics features was able to classify 75.1% of the 2167 CT nodule images (up to 5 images/CT slices per nodule) and 80% of the nodules correctly. The best of the Deep Learning networks achieved almost equivalent results.

      The image classification deep neural network results suggest the implementation of more advanced regularization and initialization deep learning techniques to further refine the decision boundaries for these pretrained networks might be benefitial. We believe the development of visualization neural network software to highlight the defining nodule features during classification would clinically assist in providing context clues for nodule diagnosis.

      This work has been supported by TFRI project ref:1068

      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.