Multiscale convolution block U-Net for automatic epidermis segmentation in immunofluorescence images
Journal
Biomedical Signal Processing and Control
Journal Volume
109
Start Page
108027
ISSN
1746-8094
Date Issued
2025-11
Author(s)
Abstract
Small fiber neuropathy (SFN) is an agonizing disease that decreases and degenerates peripheral nerve fibers in the skin. Intraepidermal nerve fiber density examination plays an essential role to serve as a decidedly sensitive indicator of sensory nerve fiber damage pertinent to the SFN-related diseases. High-resolution immunofluorescence microscopy is a contemporary imaging technique that provides sharper visibility of the tissue structures. Segmentation of the epidermis region in immunofluorescence images is a prior image processing task, which is critical to the understanding of the diseases. Due to the deficiency of appropriate tools for this particular mission, this paper explores the possibility of an automatic epidermis segmentation framework in skin immunofluorescence images. To preserve the high resolution while increasing the image amount, each original image is partitioned into a series of subimages with a manageable dimension for model processing. A U-shaped network structure consisting of five layer blocks with different depths is investigated for epidermis segmentation. The central unit is the multiscale convolution block, which comprises two different convolution pipelines. By integrating and participating the information from these multiscale convolution blocks with intensive skip connections between them, we can effectively strengthen the diversity of the feature maps for more accurate segmentation. The proposed epidermis segmentation network was evaluated on an in-house epidermis image dataset in comparison with seven state-of-the-art deep learning-based methods. Experimental results indicated that our segmentation system achieved the largest overall Dice score of 92.80%, which outperformed other competing network models. It was believed that this new segmentation scheme is promising in facilitating the epidermis extraction task for further SFN-related investigation.
Subjects
Deep learning
Epidermis
Image segmentation
Immunofluorescence
Small fiber neuropathy
Publisher
Elsevier BV
Type
journal article