A Multi-Backbone Pipeline Combining Biomedical Foundation Models and CNNs for Explainable Triage of Cutaneous Ulcers

Distinguishing vascular from non-vascular cutaneous ulcers is a critical clinical task that drives treatment selection. However, in non-specialist settings where most chronic ulcers are first assessed, misdiagnosis rates remain alarmingly high. This paper describes the design of an explainable binary classification pipeline that addresses this gap by combining three architecturally distinct backbones through late fusion: EfficientNet-B0 (a convolutional neural network), BiomedCLIP (a biomedical vision–language Vision Transformer), and UNI (a computational pathology Vision Transformer). The proposed fusion configuration aims to evaluate the individual and combined contributions of each backbone systematically. The pipeline incorporates Macenko colour normalisation to handle photographic variability, and dual explainability (GradCAM++ for the CNN and attention rollout for the Vision Transformers) to support clinical interpretability. The architectural rationale and the training protocol are presented and discussed for the intended clinical deployment as a triage tool in primary care and telemedicine.