AI-assisted body-region recognition for localized scleroderma: a foundational step towards automated LoSCAT scoring

Background Localized scleroderma (LS) is a chronic inflammatory skin disease that can cause long-term functional and cosmetic morbidity, particularly in children. The Localized Scleroderma Cutaneous Assessment Tool (LoSCAT) is the standard for clinical evaluation but remains manual and time-intensive. Automating components of LoSCAT could improve scalability and consistency for clinical monitoring and research. Methods We retrospectively curated the largest number of de-identified clinical photographs from pediatric LS patients enrolled in the National Registry for Childhood Onset Scleroderma during 2010–2025. We developed a privacy-preserving R Shiny platform to support image labeling and quality control using LoSCAT-defined anatomic regions. Six trained raters labeled 5,965 images. For model development, we retained only single-label images and merged the original 18 LoSCAT-defined body regions into 12 side-agnostic classes. After this filtering process, a total of 2,145 high-quality images were included for training and evaluation. A Vision Transformer (ViT) model was fine-tuned using a patient-level stratified group split (training: 1,790 images; test: 355 images) to prevent leakage and preserve class distribution. Results The ViT classifier achieved 0.805 accuracy and 0.776 macro-F1 on the held-out test set. Misclassifications occurred primarily between adjacent or visually similar regions (e.g., abdomen vs lower back; thigh vs leg), consistent with clinically plausible anatomic ambiguity rather than random errors. The trained classifier was subsequently applied to additional images to generate body-region labels at scale. Conclusion Automated body-region recognition in pediatric localized scleroderma photographs is feasible and clinically interpretable. This work provides a reproducible, privacy-preserving workflow that enables large-scale image organization and establishes a foundation for future region-specific lesion scoring. This study focuses on body-region classification as a necessary and foundational step toward automated LoSCAT scoring