Segmentation of 3D OCT Images of Human Skin Using Neural Networks with U-Net Architecture

The aim of the study is a comparative analysis of algorithms for segmentation of three-dimensional OCT images of human skin using neural networks based on U-Net architecture when training the model on two-dimensional and three-dimensional data.

Two U-Net-based network architectures for segmentation of 3D OCT skin images are proposed in this work, in which 2D and 3D blocks of 3D images serve as input data. Training was performed on thick skin OCT images acquired from 7 healthy volunteers. For training, the OCT images were semi-automatically segmented by experts in OCT and dermatology. The Sørensen-Dice coefficient, which was calculated from the segmentation results of images that did not participate in the training of the networks, was used to assess the quality of segmentation. Additional testing of the networks' capabilities in determining skin layer thicknesses was performed on an independent dataset from 8 healthy volunteers.

In evaluating the segmentation quality, the values of the Sørensen-Dice coefficient for the upper stratum corneum, ordered stratum corneum, epidermal cellular layer, and dermis were 0.90, 0.94, 0.89, and 0.99, respectively, for training on two-dimensional data and 0.89, 0.94, 0.87, and 0.98 for training on three-dimensional data. The values obtained for the dermis are in good agreement with the results of other works using networks based on the U-Net architecture. The thicknesses of the ordered stratum corneum and epidermal cellular layer were 153±24 and 137±17 μm, respectively, when the network was trained on two-dimensional data and 163±19 and 137±20 μm when trained on three-dimensional data.

Neural networks based on U-Net architecture allow segmentation of skin layers on OCT images with high accuracy, which makes these networks promising for obtaining valuable diagnostic information in dermatology and cosmetology, e.g., for estimating the thickness of skin layers.