In 2019, Google AI scientists along with radiologists released a deep learning end-to-end model for classifying CT scan volumes aimed at developing a Computer-Aided Diagnosis (CAD) tool [1]. This model generated a lot of excitement in the biomedical community, and although the Google team demonstrated promising results, there are several limitations that have been noted in the literature, including the need for further validation [2]. Our goal is to:
(1) explore the only third-party public implementation of the Google model [3] and perform our own re-implementation if necessary (e.g., if the implementation of [3] achieves different results from the original model).
(2) explore the current methods for attacking object detection models, and quantify the model’s susceptibility to adversarial examples. We will also develop new methods for generating adversarial examples for detection models working with 3D images and for identifying potential ways of defending against such attacks. We will thus investigate whether end-to-end models trained on CT scan volumes suffer from the same vulnerabilities as models trained on natural images.

[1] Ardila, D., Kiraly, A.P., Bharadwaj, S. et al. End-to-end lung cancer screening with three-dimensional deep learning on low-dose chest computed tomography. Nat Med 25, 954–961 (2019). https://doi.org/10.1038/s41591-019-0447-x
[2] Jacobs, C., van Ginneken, B. Google’s lung cancer AI: a promising tool that needs further validation. Nat Rev Clin Oncol 16, 532–533 (2019). https://doi.org/10.1038/s41571-019-0248-7
[3] Daniel Korat. 3D Neural Network for Lung Cancer RiskPrediction on CT Volumes, July 2020. https://github.com/danielkorat/Lung-Cancer-Risk-Prediction

Garrett T. Kenyon Ph.D., Los Alamos National Laboratory