Lesion-aware ordinal consistency learning for interpretable facial acne severity assessment

Accurate facial acne severity assessment is important for image-based disease monitoring and dermatological decision support. Existing automated methods usually formulate acne analysis as either lesion detection or image-level classification, limiting their ability to jointly capture local lesion evidence, ordered severity structure, and interpretable grading cues. We propose a lesion-aware ordinal consistency learning framework for facial acne severity assessment from high-resolution images. The framework integrates a Swin Transformer backbone, BiFPN multi-scale fusion, an FCOS lesion detection branch, and a CORAL ordinal severity branch. A lesion–severity consistency strategy aligns lesion-derived severity evidence with image-level ordinal predictions. Experiments on AcneSCU show that the proposed model improves both lesion localization and severity grading compared with detection-only, classification-based, ordinal-regression, cascade, dermatology foundation-model, and multi-task baselines. Additional validation on ACNE04 and other external acne image sources shows that zero-shot performance decreases under acquisition-domain shift but can be recovered through target-domain fine-tuning. Ablation, lesion-weight sensitivity, calibration, efficiency, and failure-case analyses further characterize the robustness and limitations of the framework. These results suggest that linking lesion-level evidence with ordinal severity prediction provides a transparent computational approach for facial acne assessment, while prospective validation across diverse populations and imaging devices remains necessary before clinical deployment.