# Analysis of injured-skin SS-OCT images based on combined attention UNet

## TL;DR

This study uses improved UNet models with attention mechanisms to analyze laser-induced skin damage in mice using SS-OCT images, showing strong correlations between laser dose, recovery time, and damage volume.

## Contribution

The novel contribution is the development and evaluation of three attention-based UNet models for accurate segmentation and quantitative analysis of laser-induced skin damage in SS-OCT images.

## Key findings

- ParallelAT-UNet achieved a Dice coefficient of 0.9364 and 99.39% accuracy in segmenting skin damage regions.
- Laser doses between 44.2 J/cm² and 74.4 J/cm² caused significant changes in skin damage volume, varying with dose and recovery time.
- All groups showed healing by 14 days post-treatment, with damage volumes smaller than initial measurements.

## Abstract

Optical coherence tomography (OCT) is a noninvasive imaging technique that provides high-resolution images of superficial skin tissues and has become widely used for diagnosing various skin disorders. Assessing laser-induced skin tissue damage is essential for understanding the healing mechanisms and optimizing treatment strategies. However, effectively quantifying skin damage and its correlation with laser dosage and recovery time poses a challenge. In this study, we established a laser-induced skin injury model in mice, utilizing 1 μm–2 μm laser wavelengths. We obtained SS-OCT images of the injury site under different laser doses and recovery times. To enhance image clarity, we applied noise reduction using the BM3D algorithm. We employed an improved UNet network model that incorporates SimAM and PSA modules, forming three attention mechanisms: TandemAT-UNet, ParallelAT-UNet, and NestedAT-UNet. These models were used to segment the damaged skin regions, followed by a 3D reconstruction method to quantitatively evaluate the volume of skin damage while analyzing changes about laser dose and recovery time.The BM3D algorithm effectively suppressed high-noise components, significantly improving image clarity. Among the three models, ParallelAT-UNet exhibited the best segmentation performance, achieving a Dice coefficient of 0.9364, mean Pixel Accuracy (mPA) of 92.67%, mean Intersection over Union (mIoU) of 96.31%, and an accuracy of 99.39%. Quantitative analysis revealed that laser doses between 25.0 J/cm2 and 36.5 J/cm2 caused minimal changes in skin damage volume, while doses ranging from 44.2 J/cm2 to 74.4 J/cm2 resulted in significant changes, which varied according to both the dose and recovery time. All groups showed signs of healing by 14 days post-laser treatment, with damage volumes smaller than the initial values. This study presents an efficient and reliable method for the quantitative assessment of laser-induced skin damage using OCT imaging. The findings demonstrate a strong relationship between laser dosage, recovery time, and skin damage, highlighting potential applications for noninvasive diagnosis and treatment monitoring using OCT.