Frequency Filtering for Data Augmentation in X-Ray Image Classification

Sang-Chul Lee, Ju-Hyeon Nam

Abstract:

A sufficiently large dataset is needed to solve tasks based
on deep learning. However, using medical datasets for deep
learning is challenging because their access is limited. In
this paper, we propose a novel data augmentation paradigm
that can be used in image classification of convolutional neural network-based approaches. As the frequency domain has
more unique patterns than the spatial domain, more diverse
patterns can be obtained by arbitrarily changing the frequency
domain patterns of the given image. First, we assume that
the meaningful patterns in the frequency domain are typically distributed in high-intensity regions in the Fourier spectrum. Next, we select the angle to reject meaningful patterns
from the Fourier spectrum. Subsequently, we generate a novel
mask pattern and remove specific frequency patterns in a fast
Fourier transform image. Thereafter, we apply an inverse fast
Fourier transform to the rejected frequency image to convert
it back to the spatial domain. Our method achieves consistent
performance improvements on X-ray images using various
backbones. On average, we achieved 3.30%, 7.09%, 7.75%,
8.14%, and 4.91% improvements on the accuracy, precision,
recall, F1-score, and AUC over the previous methods.