Rotation-Equivariant Semantic Instance Segmentation on Biomedical Images

Advances in image segmentation techniques, brought by convolutional neural network (CNN) architectures like U-Net, show promise for tasks such as automated cancer screening. Recently, these methods have been extended to detect different instances of the same class, which could be used to, for example, characterize individual cells in whole-slide images. Still, the amount of data needed and the number of parameters in the network are substantial. To alleviate these problems, we modify a method of semantic instance segmentation to also enforce equivariance to the p4 symmetry group of 90-degree rotations and translations. We perform four experiments on a synthetic dataset of scattered sticks and a subset of the Kaggle 2018 Data Science Bowl, the BBBC038 dataset, consisting of segmented nuclei images. Results indicate that the rotation-equivariant architecture yields similar accuracy as a baseline architecture. Furthermore, we observe that the rotation-equivariant architecture converges faster than the baseline. This is a promising step towards reducing the training time during development of methods based on deep learning.