Learning-based local quality assessment of reflectance confocal microscopy images for dermatology applications

Background and objective Skin cancer is one of the most common types of cancer and its early diagnosis significantly reduces patient morbidity and mortality. Reflectance confocal microscopy (RCM) is a modern and non-invasive method of diagnosis that is becoming popular amongst clinical dermatologists. The frequent occurrence of artifacts in the images is one of the most challenging factors in making a diagnosis based on RCM. It impedes the diagnosis process for the dermatologist and makes its automation difficult. In this work, we employ artificial neural networks to propose a local quality assessment system. It allows for the detection of artifacts and non-informative component images both retrospectively or in real-time during the examination. Methods In this research we address the quality assessment issue by proposing an artificial intelligence-based solution. 612 RCM mosaics were divided into small component images and manually classified in order to train the ResNeXt model in the quality verification context. A trained network was used to create an application that marks individual classes of the component images on the mosaic. Results We achieved the average classification precision of 0.98 both for the validation and test data sets. In addition, we present local quality assessment statistics of the 1540 cases of skin lesions to show which types of skin lesions most often present with artifacts in their RCM images. Conclusions In this research we investigate the utility of the deep convolution neural networks for the local quality assessment of the RCM images. We propose an AI-based system that may be effectively used as real-time support for the dermatologist during a RCM examination and as a base for the automation of the diagnostic process