Anatomical Skin Segmentation in Reflectance Confocal Microscopy with Weak Labels

Abstract

Reflectance confocal microscopy (RCM) allows in-vivo microscopic examination of human skin and is emerging as a powerful tool for a wide range of dermatological problems. Clinical use of RCM is limited by the need for trained experts to interpret images and the lack of supporting tools for quantitative evaluation of the images, especially in large datasets. The first task in understanding RCM images is to understand and produce a segmentation of the anatomical strata of human skin. This work presents such an algorithm using only weak supervision, in the form of labels for whole en-face sections, to learn a per pixel segmentation of a complete RCM depth stack. Using a bag-of- features representation for image appearance, and a conditional random field model for strata labels through the depth of the skin, a structured support vector machine was trained to label individual pixels. The approach was developed and tested on a dataset of 308 depth stacks from 54 volunteers, consisting of 16,144 total en-face sections. This approach accurately classified 85.7% of sections in the test set, and was able to detect underlying changes in the skin strata thickness with age.