Multimodal Deep Learning, Longitudinal Severity Tracking, and Explainable AI in Canine Dermatology: A Review of Emerging Paradigms and Decision-Support Systems

This review examines emerging applications of deep learning, multimodal data fusion, explainable artificial intelligence (XAI), and longitudinal severity tracking in canine dermatology. It critically synthesizes research published between 2021 and 2026, covering convolutional neural networks, EfficientNet, YOLO-based lesion detection, MobileNetV2, Vision Transformers, and the integration of clinical metadata with dermatological images. The review discusses the limitations of visual-only classification, including duplicate-image leakage, limited dataset diversity, insufficient external validation, and uncertainty in real-world clinical performance. It also examines explainability methods such as Gradient-weighted Class Activation Mapping (Grad-CAM) and SHAP, alongside the challenges of tracking lesion severity across repeated examinations. The paper proposes the PetZoe VetScan paradigm: a veterinarian-in-the-loop decision-support architecture combining lesion detection, multimodal clinical metadata, visual explanations, uncertainty estimation, and longitudinal comparison. The proposed system is intended to support preliminary screening and monitoring rather than replace professional veterinary diagnosis or clinical judgment. Its longitudinal severity-tracking component is identified as experimental and requires further validation. The review highlights research priorities including independent external validation, diverse clinical datasets, patient-level data splitting, standardized severity assessment, transparent uncertainty communication, and prospective clinical evaluation. Keywords: Canine dermatology; deep learning; explainable artificial intelligence; Grad-CAM; longitudinal lesion tracking; multimodal clinical decision support; veterinary artificial intelligence; veterinarian-in-the-loop.