Transfer Learning using Denoising Auto-Encoders for Cellular-Level Annotation of Tumor in Pathology Slides
Adversarial examples can produce altered classifications using only seemingly innocuous, imperceptible perturbations to the original image. The imperceptibility of adversarial perturbations suggests that the corresponding classifiers use decision criteria different than those of a human. In a medical setting, inexplicable decision criteria confound a pathologist’s willingness to trust machine-generated annotations. Here, we analyze denoising tumor detection models to see if they are robust to imperceptible adversarial perturbations. Moreover, to be more fully trusted by pathologists, we require tumor detectors that generate interpretable annotations which segment pathology slides into tumorous and normal regions at the cellular level. We therefore compare transfer learning based on two different autoencoder architectures, one derived from a deep denoising bottleneck autoencoder and one from an over-complete sparse autoencoder. Both autoencoders were first trained in an unsupervised manner on a set of pathology slides drawn from the Camelyon16 dataset. The latent representations produced by each autoencoder were then passed to separate neural networks that were trained in a supervised manner on binary tumor-normal masks generated by pathologists at cellular resolution. Both tumor detectors supported better than 90% AUC PR as measured by the area under the precision/recall curve on a held-out pathology slide. To assess the underlying decision criteria used by both tumor detectors, we constructed imperceptible adversarial examples which reduced the AUC PR of both models to less than 70%. Random noise of the same amplitude had almost no effect on the AUC PR of either model. Additionally, each tumor detector was resistant to adversarial “transfer” attacks targeting the other. The adversarial perturbations showed strong characteristic differences: the deep denoising models perturbations were a very diffuse, seemingly unrecognizable pattern while the sparse coding models perturbations showed traces of tissue cells.