Machine Learning Analysis Of Connective Tissue Networks Enables Objective Characterization Of Skin Fibroses
Malini Chinta1, Shamik Mascharak, BA1, Mimi R. Borrelli, MBBS, MSc1, Alessandra L. Moore, MD1, Rachel E. Brewer1, Jan Sokol1, Gabriela Kania2, Evelyn Garibay1, Deshka Foster, MD1, Heather desJardins-Park, AB1, Bryan Duoto1, Oliver Distler2, Geoffrey C. Gurtner, MD1, H. Peter Lorenz, MD1, Derrick C. Wan, MD1, Howard Y. Chang, MD, PhD1, Michael T. Longaker, MD, MBA1.
1Stanford University, Palo Alto, CA, USA, 2University Hospital Zurich, Zurich, Switzerland.
PURPOSE: Clinical evaluation of dermal fibroses relies on histopathological analysis, which is inherently observer-dependent. Visual analysis is subjective and may preclude detection of subtle phenotypic changes in early-stage or less-severe disease. We present an image processing algorithm which enables objective quantification of multiple parameters of connective tissue architecture. We then classify histologic specimens by their respective dermal fibrotic pathologies, solely using machine learning analysis of their collagen networks.
METHODS: Ninety-five human specimens were obtained from the following diagnoses: normal skin, scar, striae distensae (stretch marks), hypertrophic scar, keloid, and scleroderma. Mouse dorsal skin and scar specimens were also obtained. Formalin-fixed, paraffin-embedded histologic specimens were stained with Picrosirius-Red, imaged by polarization microscopy, and analyzed using our image processing algorithm in Matlab 2017a. In brief, this algorithm employs color deconvolution, adaptive filtering, and skeletonization of individual collagen fibers followed by quantification of parameters such as fiber length, branching, and randomness. A neural network was trained on connective tissue parameters (using 70% of images), validated (15% of images), and finally tested (15% of images) on histological images of human specimens.
RESULTS: Using our image processing algorithm, 26 connective tissue parameters were identified and quantified. To validate the algorithm, mouse unwounded skin and scar specimens were compared. Using unsupervised hierarchical clustering, these specimens clustered by specimen type (normal skin vs scar) based on four clusters of fiber parameters. The algorithm was then applied to human specimens (unwounded skin, striae distensae, "normal" scars, hypertrophic scars, and keloid). These human specimens were differentiated by five parameter clusters due to the larger degree of variation in connective tissue architecture (Figure 1). The trained neural network classified pathologies with an overall accuracy of 86% (ROC curves > 95% for all specimens), demonstrating high sensitivity and specificity (Figure 2A). The neural network also differentiated normal human skin from preclinical scleroderma with a 91% overall accuracy (ROC curves > 95%), demonstrating that our algorithm detected early-stage disease prior to the onset of clinical symptoms (Figure 2B).
CONCLUSIONS: We present an automated machine learning analysis pipeline for objective characterization of dermal collagen networks. Using a trained neural network, we classify human fibrosis specimens into disease categories based on quantitative analysis of their connective tissue properties alone. The ability to objectively characterize dermal fibroses and to detect preclinical disease has significant implications for clinical diagnosis and management as well as basic research. We intend to expand the use of this technology to fibroses in both skin and other organs, with the goal of establishing a standardized approach for histologic analysis of fibrosis.
Back to 2019 Abstracts