The effect of blurring on lung cancer subtype classification accuracy of convolutional neural networks

This is the authors' abstract. We don't add key points for this paper.

Deep learning models are extensively used for analyzing hematoxylin and eosin stained whole slide images. They enjoy high prediction accuracies, but may suffer large performance drops when applied to out-of-sample data. Here we systematically investigate how resolution differences between train and test sets may affect lung cancer subtype predictions from whole slide images using a transfer learning model based on the Inception V3 network. We observe models trained on blurred images perform well when applied to test sets with similar blurrings, but suffer poor predictions when applied to images with large blurring differences. In particular, we observed low area under curve values when models trained on blurred images were applied to non-blurred images.