Answers are grounded in this paper and cite their sources.
Pediatric facial image classification remains challenging due to the limited representation of rare conditions and the heterogeneity of available image sources. This study presents FAITHAI, an integrated framework combining facial region-of-interest standardization, StyleGAN-3-based class-conditional image synthesis, DenseNet-169 classification, and explainable artificial intelligence. The nine-class FACED3 classifier was evaluated on 221 held-out real images, achieving an accuracy of 0.733 (95% Wilson confidence interval: 0.671–0.787), balanced accuracy of 0.645, macro F1-score of 0.635, and macro one-vs-rest ROC AUC of 0.955. Across seven random-seed runs, the mean accuracy was 0.732 ± 0.034. A complementary binary classification analysis using real images only was conducted on 525 held-out images, achieving an accuracy of 0.836 and ROC AUC of 0.904 for the seed-11 model, with a mean accuracy of 0.817 ± 0.016 across ten runs. Quantitative Grad-CAM analysis on the nine-class test set yielded a deletion AUC of 0.217 compared with 0.471 for random deletion, with lower paired deletion AUC in 216 of 221 images. Attribution maps demonstrated high stability under small brightness perturbations (median correlation: 0.956) and substantial changes following classifier-head randomization (median correlation: 0.125). These tests characterize computational attribution behavior; the clinical and anatomical relevance of highlighted regions has not been established. Overall, FAITHAI provides an integrated approach to pediatric facial image classification by combining synthetic-image-assisted learning, quantitative performance evaluation, and explainable AI. The findings demonstrate its applicability to multiclass and binary facial image classification while providing complementary insights into model discrimination and attribution behavior.
No takes yet. Share an insight, caveat, or question.
Shah et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: