PT - JOURNAL ARTICLE AU - Joel, Marina Z. AU - Umrao, Sachin AU - Chang, Enoch AU - Choi, Rachel AU - Yang, Daniel AU - Duncan, James AU - Omuro, Antonio AU - Herbst, Roy AU - Krumholz, Harlan AU - Aneja, Sanjay TI - Adversarial Attack Vulnerability of Deep Learning Models for Oncologic Images AID - 10.1101/2021.01.17.21249704 DP - 2021 Jan 01 TA - medRxiv PG - 2021.01.17.21249704 4099 - http://medrxiv.org/content/early/2021/02/13/2021.01.17.21249704.short 4100 - http://medrxiv.org/content/early/2021/02/13/2021.01.17.21249704.full AB - Background Deep learning (DL) models have shown promise to automate the classification of medical images used for cancer detection. Unfortunately, recent studies have found that DL models are vulnerable to adversarial attacks, which manipulate images with small pixel-level perturbations designed to cause models to misclassify images. There is a need for better understanding of how adversarial attacks impact the predictive ability of DL models in the medical image domain.Methods We examined adversarial attacks on DL classification models separately trained on three medical imaging modalities commonly used in oncology: computed tomography (CT), mammography, and magnetic resonance imaging (MRI). We investigated how iterative adversarial training could be employed to increase model robustness against three first-order attack methods.Results On unmodified images, we achieved classification accuracies of 75.4% for CT, 76.4% accuracy for mammogram, and 93.6% for MRI. Under adversarial attack, model accuracy showed a maximum absolute decrease of 49.8% for CT, 52.9% for mammogram, 87.3% for MRI. Adversarial training caused model accuracy on adversarial images to increase by up to 42.9% for CT, 35.7% for mammogram, and 73.2% for MRI.Conclusion Our results indicated that DL models for oncologic images are highly sensitive to adversarial attacks, as visually imperceptible degrees of perturbation are sufficient to deceive the model the majority of the time. Adversarial training mitigated the effect of adversarial attacks on model performance but was less successful against stronger attacks. Our findings provide a useful basis for designing more robust and accurate medical DL models as well as techniques to defend models from adversarial attack.Competing Interest StatementThe authors have declared no competing interest.Funding StatementThis work was funded in part by a Career Enhancement Program Grant (PI: Aneja) from the Yale SPORE in Lung Cancer (1P50CA196530) and by a Conquer Cancer Career Development Award (PI: Aneja), supported by Hayden Family Foundation. Any opinions, findings, and conclusions expressed in this material are those of the author(s) and do not necessarily reflect those of the American Society of Clinical Oncology or Conquer Cancer, or Hayden Family Foundation.Author DeclarationsI confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.YesThe details of the IRB/oversight body that provided approval or exemption for the research described are given below:Research was conducted in accordance with the Declaration of Helsinki guidelines and approved by the Yale University Institutional Review Board (Protocol ID: HIC#2000027592). Informed consent was obtained from all participants in this study.All necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived.YesI understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).YesI have followed all appropriate research reporting guidelines and uploaded the relevant EQUATOR Network research reporting checklist(s) and other pertinent material as supplementary files, if applicable.YesAll data is available from the authors upon reasonable request.