Hybrid Rotation Invariant Networks for small sample size Deep Learning

Alexander Katzmann, Marc-Steffen Seibel, Alexander Mühlberg, Michael Sühling, Dominik Nörenberg, Stefan Maurus, Thomas Huber, Horst-Michael Groß

Keywords: rotational invariance, regularization, colorectal cancer, pancreatic cancer, liver lesion segmentation

TL;DR: A method for inherently invariant convolutional neural networks for small sample size deep learning and its application on medical imaging data.

Abstract: Medical image analysis using deep learning has become a topic of steadily growing interest. While model capacity is continiously increasing, limited data is still a major issue for deep learning in medical imaging. Virtually all past approaches work with a high amount of regularization as well as systematic data augmentation. In explorative tasks realistic data augmentation with affine transformations may not always be possible, which prevents models from effective generalization. Within this paper, we propose inherently rotationally invariant convolutional layers enabling the model to develop invariant features from limited training data. Our approach outperforms classical convolutions on the CIFAR-10, CIFAR-100, and STL-10 datasets. We show the transferability to clinical scenarios by applying our approach on oncologic tasks for metastatic colorectal cancer treatment assessment and liver lesion segmentation in pancreatic cancer patients.

Code Of Conduct: I have read and accept the code of conduct.