Background/Purpose: Localized scleroderma (LS) is a chronic inflammatory skin disorder that can cause functional and cosmetic impairment, especially in children. The Localized Scleroderma Cutaneous Assessment Tool (LoSCAT) is the current standard for clinical evaluation but relies on manual and time-consuming visual scoring. With recent advances in artificial intelligence (AI), there is growing interest in automating components of this process to improve efficiency.

Methods: Clinical photograph images of pediatric LS patients were retrospectively collected at the Children’s Hospital of Pittsburgh between 2010 and 2025. A custom R Shiny platform was developed with two interfaces to support data labeling and quality control, which is Grade Images and Results Summary. Six trained raters annotated 5,965 de-identified clinical images according to the 18 LoSCAT-defined body regions, later merged into 12 side-agnostic classes to improve model robustness. A Vision Transformer (ViT) model (vit_base_patch16_384) was fine-tuned using stratified group splits (train: 1,790 images; validation : 355 images) with balanced sampling, label smoothing, and a two-phase fine-tuning strategy.

Results: The ViT-based classifier achieved a macro-F1 score of 0.776 on the validation set, demonstrating reliable anatomical region annotation. The confusion matrix demonstrated that most predictions of anatomic areas were correctly identified, and the majority of errors occurred between neighboring or visually similar anatomical regions, such as the abdomen and lower back or the thigh and leg. This pattern suggests that the model effectively learned the spatial structure of the human body rather than producing random misclassifications. When applied to the remaining unlabeled dataset, the classifier successfully assigned region labels to all images, enabling linkage with existing LoSCAT clinical scores for downstream analysis of lesion distribution and disease activity.

Conclusion: This work demonstrates the feasibility of using deep learning to automate body-region identification in localized scleroderma images. The next step is to train the model with the components of the LoSCAT paired with clinical scoring. The proposed system enables large-scale image organization and provides a foundation for future region-specific lesion scoring and automated LoSCAT computation, advancing toward more objective and efficient clinical monitoring of LS.

« Back to 2026 Pediatric Rheumatology Symposium

ACR Meeting Abstracts - https://acrabstracts.org/abstract/ai-assisted-body-region-recognition-for-localized-scleroderma-the-1st-step-in-automated-loscat-scoring/