Pre-processing dermatological image data to significantly improve the performance of cyclic GAN models

Artificial Intelligence (AI) has found its place within the field of medical diagnostics. It can assist in feature recognition, in image segmentation and enhancement or in assessing specific structures. In the field of dermatology, there is a special interest in AI supported tools to classify skin lesions. However, in order to train such a classifier, lots of image data is needed. The available image data on skin lesions shows a large variety in size, resolution and lighting as there is no underlying acquisition standard. These properties make up the main components of the image data but do not contain any information on the actual subject and hence do hinder the training process. In this work, a method for cleaning image data is presented. It aims to standardize the image data in the aspects of image size, section and lighting without lowering the resolution and while maintaining the relevant structures shown. To achieve this, the images are framed to ensure they all have the same size. In the next step, microscopic images are selected and their region of interest is determined. The lighting is adjusted by standardizing the colors. Afterwards, the final size adjustment follows. The results were evaluated by carrying out a t-distributed stochastic neighbor embedding as well as a Principal Component Analysis. It could be shown that the presented routine improves the separability of benign and malign skin lesions which can also speed up and increase the quality of training process of AI models.