cs.CVSep 25, 2026
SaveCG-HAF: An Interpretable Global-Local Lesion-Burden Fusion Framework for Ordinal Acne Severity Grading in Agentic Skincare Support
Organizations: Department of Computer Science, International Islamic University Chittagong (IIUC), Chittagong, Bangladesh · Department of Mechatronics, Rajshahi University of Engineering and Technology (RUET), Rajshahi, Bangladesh · Department of Computer Science and Engineering, College of Applied Studies, King Saud University, Riyadh, 11437, Saudi Arabia · AI & Digital Health Technology, Rural Health Research Institute, Charles Sturt University, Orange, NSW, 2800, Australia · AI & Digital Health Technology, Artificial Intelligence and Cyber Futures Institute, Charles Sturt University, Bathurst, NSW, 2795, Australia · School of Health and Rehabilitation Sciences, The University of Queensland, St Lucia, Brisbane, QLD, 4072, Australia
Abstract
Ordinal acne severity grading requires distinguishing visually similar neighboring grades while jointly weighing holistic facial appearance and localized lesion burden - evidence that most existing approaches collapse into a single opaque representation. We introduce CG-HAF, a global-local fusion framework that instead keeps this evidence explicit: averaged holistic severity probabilities from independently trained classifiers are combined with structured lesion-burden descriptors from an object detector (lesion count, detection confidence, lesion area) into a compact representation, from which a lightweight, interpretable classifier produces the final grade. On a widely used benchmark, this fusion yields a clear, statistically supported improvement over global-evidence-only baselines, with the largest gains on the most severe cases. Testing on an independent dataset with a different grading standard shows that strong within-dataset performance does not transfer automatically, and a follow-up diagnostic attributes much of this gap to mismatched grading criteria rather than detection failure alone. These findings support interpretable global-local fusion as an effective strategy for ordinal acne grading while highlighting criterion alignment as key to cross-dataset portability, with a further illustration of how the resulting severity signal can support transparent, non-diagnostic decision-making in skincare applications.
Figures & tables
| Study | Dataset / task | Core approach | Evidence / fusion strategy | Distinction from CG-HAF |
| Zhang and Ma (2022) [ 15 ] | ACNE04; severity, count, localization | Joint classification–localization ensemble | Severity/count prediction with parallel localization outputs | Parallel outputs; detector burden not fused into final grading |
| Khairani and Kosala (2025) [ 16 ] | Acne detection; 4-class severity | YOLOv11 with count-to-grade mapping | Lesion counts converted to severity via count thresholds | Count-derived grading; no independent global probability stream |
| ClearFace (2025) [ 17 ] | Lesion detection/classification; IGA scoring | Detection, lesion-type classification, weighted scoring | Local detections/counts aggregated into an IGA score | Explicit scoring; no independent global probability stream |
| Zhang et al. (2024) [ 19 ] | ACNE04; lesion detection | Enhanced YOLOv7 detector | Multi-scale detector features for localization | Optimizes detection, not global–local grading fusion |
| Ilahi and Gunawan (2025) [ 20 ] | Multi-type lesion detection | YOLOv8 with optimizer-focused evaluation | Detector-internal boxes, classes, confidence | Detector optimization; no fusion with global severity |
| Nainggolan et al. (2025) [ 21 ] | Inflammatory lesion detection; skin-color analysis | YOLOv11 inflammatory detection | Local detection evaluated across skin-color scenarios | Detection fairness study, not final-stage grading fusion |
Table 1: Representative acne-severity grading and lesion-aware approaches positioned relative to CG-HAF. Comparisons are methodological; differences in datasets, grading criteria, and evaluation protocols preclude direct ranking of performance.
| Split | Mild | Mod. | Sev. | V. Sev. | Images | Boxes |
| Train | 362 | 457 | 139 | 91 | 1049 | 13611 |
| Validation | 32 | 59 | 13 | 12 | 116 | 1616 |
| Test | 99 | 131 | 36 | 26 | 292 | 3756 |
| Total | 493 | 647 | 188 | 129 | 1457 | 18983 |
Table 2: Fixed ACNE04 severity and lesion-annotation composition.
Figure 1: Retained CG-HAF architecture. Three-seed ConvNeXtV2-Large probabilities provide 4-D global severity evidence, while YOLOv12s detections provide 8-D structured lesion-burden evidence. Their concatenation yields the 12-D F4 representation used by a validation-fitted balanced multinomial logistic regression classifier. The auxiliary count head is training-only, and Grad-CAM/YOLO visualizations are post hoc inspection components that do not affect predictions.
Figure 2: Detailed architecture of the global ConvNeXtV2-Large severity-count branch, illustrating the backbone representation learning, four-class ordinal severity head, auxiliary 65-bin lesion-count distribution head, and Teacher-EMA consistency regularization pathway.
Figure 3: Detailed architecture of the local YOLOv12s lesion-evidence branch, illustrating localized multi-scale acne lesion detection, bounding-box prediction, detection-confidence extraction, and count-to-grade severity mapping.
| Family | ID | Controlled change | Purpose |
| Training a | Full | Reference branch: masking + CE/LDL + count guidance + Teacher-EMA | Reference configuration |
| A0 | Vanilla CE-only classifier | Effect of full training design vs. minimal baseline | |
| A1 | Remove skin-region masking | Contribution of masking | |
| A2 | Remove LDL | Contribution of ordinal label-distribution supervision | |
| A3 | Remove count head/losses | Contribution of count guidance | |
| A4 | Remove Teacher-EMA consistency | Contribution of EMA consistency regularization |
Table 3: Controlled training and fusion ablation design.
Figure 4: Implemented SkinAgent application example showing input validation, CG-HAF-derived severity and lesion-burden evidence, and downstream safety-grounded routing. This example demonstrates system integration only and is not a quantitative or clinical validation of the complete agentic system.
Figure 5: Confusion matrix of the retained three-seed probability-aggregated F4 CG-HAF on the fixed ACNE04 test split. Integer counts are shown; 36 of 37 errors occurred between adjacent severity grades.
Figure 6: Qualitative comparison of representative ACNE04 test cases using principal internal comparison systems with aligned per-image predictions. Each row shows the input image, ground-truth severity grade, and predictions from the YOLOv12s count-only reference, the historical canonical CG-HAF system, and the retained three-seed probability-aggregated F4 CG-HAF model. The examples illustrate image-level differences among the compared predictors and include a challenging case to avoid presenting only successful final-model predictions.
| Predictor | Accuracy | Macro-F1 | QWK | Notes |
| Seed 42 | 84.93% | 0.8221 | 0.9018 | Independent F4 stack |
| Seed 123 | 86.30% | 0.8441 | 0.9128 | Independent F4 stack |
| Seed 2026 | 85.27% | 0.8445 | 0.9064 | Independent F4 stack |
| Seed mean SD | 85.50% 0.71 pp | 0.8369 0.0128 | 0.9070 0.0055 | Arithmetic run-level summary |
| Probability-aggregated F4 | 87.33% | 0.8576 | 0.9173 | Final image-wise probability aggregation |
Table 4: Seed-level stability and probability-level aggregation.
Figure 7: Seed-level F4 stacking performance and the distinct probability-aggregated F4 predictor. The seed mean is a descriptive run-level summary; the aggregated predictor has no run-level error bar.
| System | Accuracy | Macro-F1 | QWK | Grade MAE | ECE | Severe recall | Very Severe recall | Decision |
| Original-branch F4 | 87.33% | 0.8576 | 0.9173 | 0.1301 | 0.1453 | 0.8333 | 0.8846 | Retained |
| Count-conditioned F4 | 86.99% | 0.8557 | 0.9214 | 0.1301 | 0.1363 | 0.8056 | 0.8846 | Secondary |
| Frozen 0.625/0.375 dual stack | 87.33% | 0.8486 | 0.9255 | 0.1267 | 0.1354 | 0.7500 | 0.9231 | Rejected |
Table 5: Safety-constrained comparison of retained and secondary three-seed candidate systems.
Figure 8: Complementary interpretability evidence for CG-HAF. The upper panels show historical qualitative ConvNeXtV2 Grad-CAM and YOLOv12s examples for aligned Severe ( levle3_21 ) and Very Severe ( levle3_16 ) images; they are not three-seed ensemble attributions. The lower panel shows coefficients from the final retained validation-fitted F4 logistic-regression stack. Coefficients are descriptive, scale-dependent, and non-causal.
| Model | Acc. | BAcc. | F1 | QWK | MAE |
| Global | 35.0 | 0.3346 | 0.3301 | 0.4390 | 0.770 |
| Local | 40.0 | 0.3532 | 0.3362 | 0.4367 | 0.730 |
| F4 | 39.0 | 0.3554 | 0.3459 | 0.4651 | 0.715 |
Table 6: Frozen subject-level cross-criterion evaluation on PLSBRACNE01.
| Method | Acc. | Macro-F1 | QWK | MAE |
| Hayashi rule | 43.5 | 0.3629 | 0.5344 | 0.605 |
| TOTAL2 ordinal | 58.5 | 0.5874 | 0.7312 | 0.475 |
| TYPE10 multinomial | 72.0 | 0.7227 | 0.8614 | 0.290 |
| TYPE10 ordinal | 74.0 | 0.7364 | 0.8726 | 0.270 |
Table 7: Annotation-assisted cross-criterion diagnostic on PLSBRACNE01. Expert lesion evidence is used in this analysis.
| Configuration | Dim. | Accuracy | Balanced Acc. | Macro-F1 | QWK | Grade MAE | Severe recall | Very Severe recall |
| Global-only | 4 | 81.16% | 0.7747 | 0.7658 | 0.8706 | 0.1952 | 0.6389 | 0.7692 |
| Local-only | 8 | 80.48% | 0.7906 | 0.7820 | 0.8853 | 0.1952 | 0.6389 | 0.8462 |
| Full CG-HAF | 12 | 87.33% | 0.8699 | 0.8576 | 0.9173 | 0.1301 | 0.8333 | 0.8846 |
Table 8: Controlled global–local branch ablation on the fixed ACNE04 test partition.
Figure 9: Qualitative analysis of global–local prediction behavior on the ACNE04 test set. Panel A shows five illustrative test images with ground-truth grades and the four-class predictive distributions from the Global-only, Local-only, and Full CG-HAF configurations (cell intensity = class probability; thick outline = predicted class), including complementary single-stream successes, one Full-only-correct case, and a fusion failure case. Panel B reports correctness-pattern frequencies across the full 292-image test set. M/Mo/S/VS: Mild/Moderate/Severe/Very Severe; G/L/F: Global-only/Local-only/Full CG-HAF.
Figure 10: Progressive F0–F4 fusion-feature ablation shown as improvement relative to F0. The confirmatory comparison concerns the complete F4 representation versus F0 (Holm-adjusted exact McNemar ); intermediate numerical increments are not interpreted as independently significant.
| Variant | Feature composition | Dim. | Accuracy | Macro-F1 | QWK | Grade MAE | Severe recall | Very Severe recall |
| RAW_CNN | Direct three-seed mean CNN | 4 | 81.16% | 0.7586 | 0.8689 | 0.1952 | 0.6111 | 0.7308 |
| F0 | Averaged global probabilities | 4 | 81.16% | 0.7658 | 0.8706 | 0.1952 | 0.6389 | 0.7692 |
| F1 | F0 + count-derived grade | 8 | 85.27% | 0.8202 | 0.9053 | 0.1507 | 0.7222 | 0.8462 |
| F2 | F1 + continuous count | 9 | 85.62% | 0.8284 | 0.9068 | 0.1473 | 0.7500 | 0.8462 |
| F3 | F2 + confidence evidence | 11 | 85.96% | 0.8367 | 0.9083 | 0.1438 | 0.7778 | 0.8462 |
| F4 | F3 + lesion area; complete 12-D | 12 | 87.33% | 0.8576 | 0.9173 | 0.1301 | 0.8333 | 0.8846 |
Table 9: Progressive decomposition of the CG-HAF detector-to-grading representation.
| Configuration | Accuracy | Balanced Acc. | Macro-F1 | QWK | Grade MAE | Severe recall | Very Severe recall |
| Full image branch | 81.16% | 0.7619 | 0.7586 | 0.8689 | 0.1952 | 0.6111 | 0.7308 |
| A0: vanilla CE-only | 80.14% | 0.7774 | 0.7792 | 0.8756 | 0.1986 | 0.6667 | 0.8077 |
| A1: without skin-region masking | 81.51% | 0.7922 | 0.7870 | 0.8884 | 0.1849 | 0.6111 | 0.8846 |
| A2: without LDL | 78.42% | 0.7241 | 0.7247 | 0.8498 | 0.2226 | 0.5000 | 0.7308 |
| A3: without count guidance | 80.48% | 0.7610 | 0.7593 | 0.8643 | 0.2021 | 0.5556 | 0.8077 |
| A4: without Teacher-EMA consistency | 79.79% | 0.7518 | 0.7522 | 0.8583 | 0.2089 | 0.6111 | 0.7308 |
Table 10: Controlled training-component ablation using three-seed probability-mean global predictions.
Figure 11: Differences between A0–A4 and the Full global image branch at the probability-mean endpoint. The plot is descriptive; paired bootstrap confidence intervals supporting the A2 finding are reported in the text and are not inferred for all variants.
| Study | Dataset / task | Reported result |
| Wu et al. [ 1 ] | ACNE04; 4-grade severity | Acc.: 84.11% |
| KIEGLFN [ 4 ] | ACNE04 / PLSBRACNE01; grading | Acc.: 84.52% / 59.35% |
| DED [ 5 ] | ACNE04 / PLSBRACNE01; grading | Acc.: 86.06% / 67.56% |
| AcneGrader [ 9 ] | ACNE04; severity grading | Acc.: 85.82% |
| Prior-Knowledge Guided [ 10 ] | ACNE04 / PLSBRACNE01; grading | Acc.: 85.27% / 65.85% |
| FF-PLL [ 11 ] | ACNE04 / ACNE-ECKH; grading | Acc.: 87.33% / 67.50% |
Table 11: Reported performance of representative acne-severity grading and lesion-aware studies, arranged in citation order.