Title:
Interpretable Image-level Acne Severity Grading via Efficientnet-B0 Transfer Learning and Grad-CAM: An Open, Cross-platform Reference Implementation Validated on the ACNE04 Benchmark

Authors:
Sophie Zeng, Sean Kalaycioglu*, Collin Hong, Haipeng Xie, Canada

Abstract:

Acne vulgaris affects most adolescents and a substantial proportion of adults; accurate severity grading determines treatment choice, monitoring, and trial endpoints, yet manual assessment on scales such as the Investigator’s Global Assessment or Hayashi criteria is limited by inter-rater variability and imaging conditions. We developed and evaluated an image-level four-class Hayashi-criterion acne severity classifier based on transfer learning of Image Net-pretrained EfficientNet-B0. The model was fine-tuned on the public ACNE04 benchmark (2,983 labeled images) with AdamW optimization, standard geometric and photometric augmentation, and checkpoint selection by validation macro-F1. On a held-out stratified 15 % test set the classifier achieved 93.5 % accuracy and 94.4 % macro-F1 (per-class F1 0.92–0.97), with 83 % of errors confined to adjacent grades; Cohen’s quadratic-weighted κ reached 0.956 (95 % CI [0.935, 0.973]). Bootstrap confidence intervals confirmed stable performance. Grad-CAM visualizations generated from the final convolutional block concentrated on clinically relevant facial regions (forehead, cheeks, chin). The complete pipeline is released as functionally equivalent open-source implementations in Python (PyTorch + timm) and MATLAB R2026a, including a clinician-facing inference widget and a tiered backbone fallback that guarantees runnability without specialized pretrained-weight packages. The results demonstrate that lightweight transfer learning can deliver strong, balanced, and interpretable grading performance on a public benchmark while providing a reproducible cross-platform reference for subsequent prospective and device-stratified validation required for clinical translation.

DOI: http://dx.doi.org/10.51505/ijaemr.2026.11406