Author of

• 35619

  +

  P42 - Screening and Early Detection - Risk Modelling and Artificial Intelligence (ID 177)

  • Event: WCLC 2020
  • Type: Posters
  • Track: Screening and Early Detection
  • Presentations: 1
  • Moderators:
  • Coordinates: 1/28/2021, 00:00 - 00:00, ePoster Hall

  • 35624

    +

    P42.04 - Is AI Better for Prediction of Pathologial Subtype in Pulmonary Nodules？  (ID 1713)

    00:00 - 00:00  |  Presenting Author(s): shuaibo Wang
    • Abstract

    Introduction
    

    Artificial intelligence (AI) has made great progress in the field of image recognition in recent years, especially in the location of lung nodules. Quite a few models, which are based on convolutional neural networks (CNN), have emerged and researchers have already carried out clinical trials in different countries. However, whether AI can outperform clinicians in identifying lung nodule subtypes or pathological subtypes by CT scans remains a question.

    Methods
    

    We conducted a systematic review and Bayesian diagnostic meta-analysis of retrospective studies on the comparison between artificial intelligence and doctors in the classification of pulmonary nodules. We searched Medline, Embase, Scopus, Web of science, Cochrane library, CNKI, Wangfang, VIP databases from January 2012 to July 2020. The meta-analysis was processed using the R 'bamdit' package. The gold standard for all studies was pathological diagnosis. The SROC curves were compared to evaluate the diagnostic efficacy between CNN and clinicians.

    Results
    

    A total of 13 retrospective cohort studies including 1951 patients in the comparison between AI and doctors were identified. We divided them into three tasks based on the purpose of the model. In the classification of benign and malignant nodules(task 1), the sensitivity of artificial intelligence was 0.80 (95% CI:0.58- 0.91) and the specificity was 0.75 (95%CI 0.59-0.85). However, the sensitivity and specificity of human were 0.71 (95% CI 0.63, 0.78) and 0.63 (95%CI 0.55-0.71) respectively. In the classification of PILs (pre-invasive lesions) and ILs (invasive lesions) (task 2), the sensitivity and specificity of AI were 0.69 (95% CI:0.24- 0.93) and 0.57 (95%CI 0.17-0.89), respectively, while the sensitivity and specificity of the humans were 0.79(95%CI 0.58-0.91) and 0.51(95%CI 0.35-0.68) respectively. Assessment of predictive performance in the differential diagnosis of IA/not IA (task 3) demonstrated a good fit for AI and humans with sensitivity scores of 0.83(95%CI 0.63-0.93), 0.73(95%CI 0.63-0.82) and specificity scores of 0.82(95%CI 0.66-0.92) and 0.86(95%CI 0.80-0.92) respectively.

    

    Conclusion
    

    In these thirteen retrospective studies, although AI has not completely surpassed humans, the sensitivity of AI was higher than that of the humans in task 1 and task 3, and the specificity of AI was higher than that of the humans in task 1 and task 2. However, we may need more evidences. Still, machine learning has a strong potential in the classification of lung nodules. There is an urgent need for standardized reporting guidelines in the comparison between AI and doctors. We should explore bravely, evaluate carefully, and apply prudently.