Generalization of intensity distribution of medical images using GANs

초록

The performance of CNN-based medical image classification network depends on the intensities of the trained images. Therefore, it is necessary to generalize medical images of various intensities. Previously, histogram matching was used to generalize the intensity of images, but this leads to loss of features. To maintain the performance of the medical image classification network, we propose a method using Generative adversarial network(GAN) as generator to control the arbitrary intensity distribution to the specific intensity distribution in the training set. Also, we select CycleGAN for training unpaired medical image dataset. For the performance evaluation, 1) we compare similarity of intensity distribution between generalized dataset, original dataset and histogram matched set using histogram; 2) for accuracy comparisons, the original domain dataset was passed to a classifier that only trained the original domain images. As a result, we confirmed the similarity of intensities between the generalized images and training dataset, also confirmed that AUC of our method is 0.84, which is higher than the other methods.