Machine learning to determine skin involvement in systemic sclerosis using spatial frequency domain imaging
OA Version
Citation
Abstract
Systemic Sclerosis (SSc) is a complex multi-system autoimmune disorder, which involves fibrosis of skin and internal organs. Although it is a rare disorder, SSc causes a considerable loss in quality of life and may result in death, as the result of cardiac, renal, and pulmonary complications. Skin involvement in SSc is associated with disease progression and is currently assessed using the modified Rodnan Skin Score (mRSS). The mRSS ranges from 0 (no fibrosis) to 3 (high fibrosis), and is assessed clinically by manually palpating 17 different body sites. This metric is subjective, and is prone to sampling bias, as there may be significant variation in skin involvement at a given body site. The Roblyer lab has begun to investigate Spatial Frequency Domain Imaging (SFDI), a wide-field, non-invasive imaging technique that can be used to obtain diffuse reflectance and optical properties, as an objective alternative to assess skin involvement in SSc. While preliminary studies have shown SFDI to be a promising alternative to the mRSS, past work has been limited to investigating these metrics individually and averaging these metrics over a region of interest (ROI), which may not fully encompass the variation in skin involvement in SSc. Machine learning can be used to more accurately discriminate discrete levels of skin involvement in scleroderma using a combination of SFDI derived metrics and identify the spatial distribution of skin involvement in SSc.v
Diffuse reflectance and optical properties from 8 wavelengths and 8 spatial frequencies obtained from a cohort of 25 patients and 18 controls using a commercial SFDI instrument were used as features into a Linear Discriminant Analysis model. The best performing features were identified based on ROC AUC for clinically relevant binary classifications and accuracy for multiclass classifications. Posterior probability maps were generated based on the optimal single feature classifier identified for distinguishing SSc subjects from healthy controls for the left arm. Spatial distribution of posterior probabilities and the factors affecting it were quantitatively and qualitatively assessed.
In most cases, using a combination of features significantly improved the ability to distinguish between healthy controls and the various levels of skin involvement in SSc. Multiclass classifiers showed significantly worse performance as compared to the binary classifiers. Analysis of spatial distribution showed that while the mean posterior probability of healthy controls is significantly lower than SSc subjects, healthy controls show more variation than SSc, possibly due to factors such as hair and aging. The highest separation between the mean posterior probabilities of healthy controls and SSc subjects was obtained using the smallest circular ROI, while variation stayed relatively constant across all ROI sizes, suggesting that the difference between them is localized. This study can be extended to analyzing more features such as the tissue oxygen saturation, histopathological markers, etc. and identifying the spatial distribution of other body sites using multiple features.
Description
2025