Data augmentation method for simulating lung lesion evolution

In the medical image field, deep learning, as a method of computer-aided diagnosis, is widespread and performs well, but deep learning relies on a large amount of data for training, yet it is often difficult to collect enough data for training in the medical image field. To overcome this problem, data augmentation has become a popular method to increase the size of training datasets, using data augmentation techniques to process existing samples can significantly alleviate the problem of lack of training data. At the same time, because medical images contain information about physiological structures, data augmentation needs to consider the lesion region, so that not all data augmentation methods can be applied to medical images. In this paper, we propose a data augmentation method based on the evolution phase of lung lesions for medical lung images and generate data on different evolution phases of lung lesions for training purposes while ensuring that the edge information and texture information are consistent with the lesion characteristics. Also we used our own lung dataset to experiment with our proposed method, experiment show that the data generated can ensure physiological structures.