Virtual Library

Start Your Search

William Hickes

Author of

  • +

    MA20 - Implementation of Lung Cancer Screening (ID 923)

    • Event: WCLC 2018
    • Type: Mini Oral Abstract Session
    • Track: Screening and Early Detection
    • Presentations: 1
    • Moderators:
    • Coordinates: 9/25/2018, 15:15 - 16:45, Room 206 F
    • +

      MA20.10 - Lung Cancer Prediction Using Deep Learning Software: Validation on Independent Multi-Centre Data (ID 14022)

      16:20 - 16:25  |  Author(s): William Hickes

      • Abstract
      • Presentation
      • Slides


      Artificial Intelligence software has shown promise in predicting malignancy in indeterminate CT detected pulmonary nodules. This study aimed to assess the accuracy of a convolutional neural network (CNN) based lung cancer prediction software on an independent dataset of indeterminate incidentally detected nodules in a retrospective European multicentre trial.

      a9ded1e5ce5d75814730bb4caaf49419 Method

      The software was trained using the US National Lung Screening Trial (NLST) dataset which was manually curated, such that each reported nodule and cancer was located, contoured and diagnostically characterised (9310 benign nodule patients; 1058 cancer patients). From this complete dataset, a training set was built by selecting all patients with solid and part-solid lesions of 6mm and above, where benign nodules and cancers could be confidently identified by clinicians (5972 patients, of which 575 were cancer patients). A CNN classifier was trained using Deep Learning on this data to produce a malignancy score per nodule. We defined a benign nodule rule-out test by calculating thresholds on the malignancy score that achieve 100% and 99.5% sensitivity on the NLST data.

      The study was set up so that a malignancy score for each nodule was generated. Overall performance was evaluated using Area-Under-the-ROC-Curve analysis (AUC) and rule-out performance measured the specificity at the two thresholds, i.e. the proportion of benign nodules correctly stratified at each threshold.

      There were 2201 nodules, measuring between 5-15mm from 1719 patients from three tertiary referral centres in the UK, Germany and Netherlands. The CT data included heterogeneous scan parameters, scanner manufacturers and clinical indications. Diagnostic ground-truth was established according to Fleischner or British Thoracic Society guidelines. The dataset contained 222 unique cancers from 215 patients.

      4c3880bb027f159e801041b1021e88e8 Result

      AUC on all-site data was 0.92 (95%CI = 0.89-0.93) and broken down per-site the AUC was 0.97 (Netherlands, n=883, 26 cancers), 0.93 (UK, n= 698, 51 cancers), and 0.84 (Germany, 620, 145 cancers).

      The score thresholds used for the target sensitivity of 100% and 99.5% were the same and achieved an overall sensitivity on the data of 99.1% with a specificity of 25.0%. Per-site results were 25.6% (Netherlands), 27.8% (UK) and 20.6% (Germany) specificity with 100%, 100% and 98.6% sensitivity respectively.

      8eea62084ca7e541d918e823422bd82e Conclusion

      Performance of the AI software on independent European multicentre data was comparable to that achieved on the NLST training data, although there was some variability in the performance of the system across the three centres, potentially providing an opportunity for further optimisation.


      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.