AI06 ALUDWIG: an automated artificial intelligence-based assessment of female androgenic alopecia

Abstract Female androgenic alopecia (FAGA) is a common condition affecting 3–38% of women worldwide, depending on age and population demographics (Starace M, Orlando G, Alessandrini A, Piraccini BM. Female androgenetic alopecia: an update on diagnosis and management. Am J Clin Dermatol 2020; 21: 69–84; Huang CH, Fu Y, Chi CC. Health-related quality of life, depression, and self-esteem in patients with androgenetic alopecia: a systematic review and meta-analysis. JAMA Dermatol 2021; 157: 963–70). Accurate and consistent assessment of FAGA severity is critical for diagnosis, monitoring progression and evaluating treatment efficacy. However, current severity scoring methods, such as the Ludwig scale, are subject to high intraobserver variability. These limitations underscore the need for objective, efficient, reproducible tools to standardize evaluation of FAGA across clinical and research settings. This observational study analysed 62 retrospective and 34 prospective images of FAGA collected using smartphones in a dermatology institute in Madrid. Two convolutional models were developed: one for automated head cropping and another for segmenting the hair zone and areas of hair loss. Models were trained on a larger annotated dataset from public atlases. From the model outputs, the percentage of alopecia was calculated and discretized to map it to the three-grade Ludwig scale, with model hyperparameters optimized using Bayesian techniques on retrospective images. The prospective evaluation of 34 images, conducted without tuning or hyperparameter adjustments, yielded an accuracy of 53% against investigator scores. The correlation coefficient of the model was 50% with the investigator’s alopecia percentage and 34% with the investigator’s categorical grade. This indicates that the predicted alopecia percentage aligns better with investigator scores than the categorical grade, likely due to information loss during grade conversion. A confusion matrix revealed that only two out of 34 cases were misclassified beyond one grade. This study introduces ALUDWIG, a tool for automatically computing the Ludwig score, as a promising solution for the automated assessment of FAGA severity. ALUDWIG provides a more objective and reproducible alternative to traditional scoring methods and can evaluate FAGA from a single image.