Log-Polar Image Mapping for CNN-independent Rotation and Scale Invariant Brain Tumor Classification

Abstract

Convolutional neural networks (CNNs) are widely used for image classification tasks in biomedical applications such as tumor classification. A major challenge arises when images in the test dataset differ from those used during training. While translational shifts are often handled robustly by CNNs, other perturbations such as rotation and scaling can lead to a significant decrease in performance. In this work, we apply a log-polar transformation to the input images to obtain rotation- and scale-invariant representations. We evaluate the performance across multiple CNN architectures on test data containing unseen rotational and scaling variations. The results demonstrate a significantly improved robustness to these transformations compared to the standard approach based solely on data augmentation. Furthermore, combining log-polar transformed images with data augmentation during training leads to a quasi-invariant behavior for some network architectures.