Automated Skin Biopsy Analysis with Limited Data

Abstract

In patients with diabetic and other peripheral neuropathies, the number of nerve fibers that originate in the dermis and cross the dermal-epidermal boundary is an important metric for diagnosis of early small fiber neuropathy and determination of the efficacy of interventions that promote nerve regeneration. To aid in the time-consuming and often variable process of manually counting these measurements, we propose an end-to-end fully automated method to count dermal-epidermal boundary nerve crossings. Working with images of skin biopsies immunostained to identify peripheral nerves using current standard operating procedures, we used image segmentation neural networks to distinguish between the dermis and epidermis and an edge detection neural network to identify nerves. We then applied an unsupervised clustering algorithm to identify nerve crossings, producing an automated count. Since our dataset is very small—containing less than one hundred images—we use pretrained models in combination with several image augmentation methods to improve performance on training and inference. The model learns from a human expert’s training data better than a human trained by the same expert.