Subcutaneous Tissue Structural Feature Identification Using Unsupervised Machine Learning.

Accurate quantification of the structural features of subcutaneous (SC) tissue is essential for understanding its physiology and developing predictive computational models of skin biomechanics and biotransport. Image analysis for tissue characterization, with varying feature sizes, shapes, and distribution, typically involves manual segmentation, which is labor-intensive, user-dependent, and irreproducible. Yet robust automated algorithms for these tissue structures are currently unavailable. While supervised Machine learning (ML) methods offer the potential to address this challenge, those require extensive training datasets and labeling, which are unavailable for SC tissues. In this paper, we present a new methodology that uses unsupervised machine learning to identify the structural features of subcutaneous tissue from stained histology slides. We demonstrate our method using porcine skin SC tissue samples. A novel two-dimensional (2D) image transformation generates maps of radial intensity values for each pixel, referred to as a proximal intensity map, which is subsequently reduced into a lower-dimensional feature vector space. Unsupervised learning using K-means clustering is employed to classify pixels based on their computed feature vectors. By designing the proximal intensity map and the feature space reduction, we show that the clustering step can automatically and robustly classify and identify the complex collagenous network within adipose tissue spaces. We also present an objective basis for selecting the optimal search radius based on the rationale of noise minimization from spurious segmented pixels and maximum separation of pixel intensity distribution of different features. This work provides a novel method for automatically identifying subcutaneous tissue structures, advancing the understanding of skin physiology, and developing improved in vitro tissue models.