Classificação etiológica de úlceras vasculares por modelos híbridos de deep learning explicáveis: estudo transversal

Abstract

Introduction: The etiological classification of vascular ulcers represents a significant clinical challenge due to morphological overlap among their main causes — venous, arterial, and neuropathic. Deep learning–based methods have shown promise in medical image analysis; however, concerns regarding interpretability and probabilistic reliability continue to limit their clinical adoption. Objective: To develop and evaluate an artificial intelligence model based on a hybrid CNN–Transformer architecture for the etiological classification of vascular ulcers using clinical images, integrating explainability mechanisms and probabilistic calibration. Methods: A cross-sectional study was conducted using clinical images from patients with vascular ulcers. A strict patient-level split was applied to prevent data leakage. The hybrid CNN–Transformer model was trained using transfer learning and evaluated with performance metrics (accuracy, F1-macro, and AUC-ROC), as well as calibration metrics (Expected Calibration Error and Brier Score). Interpretability was explored using Grad-CAM, SHAP, and LIME. Results: The model demonstrated clinically consistent performance in binary classification scenarios, with improved separability between more distinct etiologies. Probabilistic calibration significantly reduced the expected calibration error. Explainability analyses indicated model attention focused on morphologically relevant regions, particularly the perilesional skin in venous ulcers. Conclusion: The proposed architecture demonstrated feasibility for etiological classification based solely on clinical images, incorporating trustworthy AI mechanisms. Although limited by the absence of external validation and multimodal integration, this approach represents a methodologically rigorous step toward the translational application of artificial intelligence in vascular surgery.