Convolutional neural networks-based method for skin hydration measurements in high resolution MRI

Abstract

High-resolution Magnetic Resonance Imaging (MRI) allows the visualization of the anatomy of different skin layers and provides a wide range of physical and biochemical parameters. This technique is very effective for the study of skin hydration, through a multitude of morphological, physical and chemical properties.In this paper, we leverage the recent success achieved by deep learning architecture in several medical applications. We propose a method that segments the first layers of the skin, in order to study the hydration effect on each layer based on T2 measurements.Despite the small number of subjects studied, we were able to build a convolutional neural network (CNN) on labeled learning data and generate the T2 map to explore the effect of moisturization. Besides, CNN architecture was applied to map between the exams before and after moisturization allowing simulation of the skin hydration phenomenon.Our study proved the strong correlation between the manual measurement based on T2 mapping generation and the CNN-based measurement. The mean of the Dice index between the manual and automatic methods was 0.79 CI:[0.66-0.88] before moisturization, and 0.75 CI:[0.61-0.89] after moisturization. And the Hausdorff distance was (0.134 mm) before moisturization, and (0.226 mm) after moisturization. In our experiment, Unet was used effectively to segment skin layers, which achieved a high accuracy training score (Accuracy=0.9, Loss=0.01). For regression, an CNN model was used to simulate the skin hydration (Dice = 0.961).The Unet model used to study the hydration effect of the skin layers in 3T MRI was reliable to insure the excellent T2 measurement values according to the manual measurement and gave an ideal segmented skin layers compared to the skin anatomy. Besides, we demonstrated that CNN architecture allows simulating skin hydration.