Systematic Review of Artificial Intelligence Applications in Scar Assessment and Management

Abstract

Background: Hypertrophic and keloid scars are chronic fibroproliferative disorders with major psychosocial impact. Conventional assessment tools—Vancouver Scar Scale (VSS) and Patient and Observer Scar Assessment Scale (POSAS)—are limited by subjectivity and poor reproducibility. Artificial intelligence (AI), especially convolutional neural networks (CNNs), offers opportunities for more objective scar evaluation. Objective: To systematically review AI applications in hypertrophic and keloid scar assessment, focusing on model types, imaging inputs, performance metrics, clinical endpoints, and translational readiness. Methods: A PRISMA-compliant search of online databases identified peer-reviewed AI studies on hypertrophic or keloid scar evaluation. Non-AI, non-English, and editorial articles were excluded. Two reviewers independently screened all records, with strong inter-rater reliability (κ = 0.94). Results: Among 1520 records, 24 studies met inclusion criteria. CNNs were the most common models, followed by support vector machines (SVMs) and hybrid approaches. Imaging modalities included smartphone photography, dermoscopy, thermal imaging, second-harmonic generation microscopy, and structured light. Clinical applications involved scar classification, segmentation, recurrence prediction, and treatment monitoring. Reported performance varied widely: accuracy (63%-98.5%), sensitivity (14.9%-99.7%), specificity (80%-99.9%), AUC (0.342-1.0), Dice coefficient (0.5-0.952), and r² (0.234-0.998). Larger datasets and multimodal imaging generally improved model performance. Small or low-quality datasets produced more variable results. External validation occurred in approximately 58% of studies and typically resulted in modest performance drops, indicating overfitting. Conclusions: CNN-based models using mobile or dermoscopic imaging shows promise for objective scar assessment. Key barriers to adoption include limited external validation, explainability, and regulatory integration.