Answers are grounded in this paper and cite their sources.
Objective The aim of this study was to evaluate the diagnostic performance, calibration, and clinical interpretability of explainable deep learning models for etiological classification of lower limb vascular ulcers. Methods We conducted a cross-sectional diagnostic accuracy study using 2847 pseudonymized clinical images from 176 adult patients with venous, arterial, or neuropathic ulcers treated at a tertiary vascular surgery center. Images were independently annotated by three board-certified vascular surgeons, with consensus adjudication (Fleiss κ = 0.72). Hybrid deep learning models combining convolutional neural networks and attention-based components were trained using patient-level stratified data splits to prevent information leakage. Binary and multiclass classification tasks were evaluated. The primary outcome was macro F1 score. Model interpretability was assessed using gradient-based explainability techniques (eg, gradient-weighted class activation mapping), and probabilistic calibration was evaluated using expected calibration error. Results Binary classification performance was highest for venous vs neuropathic ulcers (accuracy, 92.4%; macro F1 score, 0.91) and venous vs arterial ulcers (accuracy, 88.3%; macro F1 score, 0.87). Multiclass classification achieved an overall accuracy of 82.1% (macro F1 score, 0.78), with most misclassifications occurring between arterial and neuropathic ulcers. Explainability analyses demonstrated concordance between model attention and clinically meaningful morphological features in 85% of evaluated cases. Probabilistic calibration improved after scaling, with expected calibration error decreasing from 0.176 to 0.049. Conclusions Explainable deep learning models demonstrated clinically coherent diagnostic performance for vascular ulcer classification, particularly in binary scenarios. These findings support the integration of interpretable artificial intelligence as a complementary decision support tool in vascular practice. Limitations include the single-center design and reliance on image-only data.
No takes yet. Share an insight, caveat, or question.
Forti et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: