Share on Facebook
Share on Twitter
Share on LinkedIn
Share by Email
Print
Automated objective grading of acne vulgaris may soon be possible, according to study results by Lim and colleagues published in Skin Research and Technology.
The study attempted to validate an algorithm developed by investigators to replace subjective analysis of the IGA tool already in use. The researchers obtained 472 facial images of varying severities of acne from 416 patients from the A*STAR Bioformatics database in Singapore (collected between 2004 and 2017) for use as training images. They coded the images into 3 IGA grades: 0-1, 2, and 3-4 (clear and almost clear, mild to moderate, and severe, respectively) to match similar treatment groups.
The study model used a convolutional neural network designed to mimic human vision to interpret the images. They were independently trained on three image sizes (reflecting 600×800, 750×1000, and 1200×1600 pixels). The automated IGA scores generated by the algorithmic model were then compared with clinical human IGA assessments for validation.
The highest accuracy of the algorithm was 67%, achieved at the highest pixel image sizes, which according to the authors, suggests that “the relatively minute features of individual acne lesions/clusters may be best captured by higher pixel densities.” Given the limitations of computer memory and the images available, they determined the optimal image size to be 1200×1600 pixels.
Comparison of the accuracy of automated evaluation at different IGA grades showed “suboptimal” performance by all models at intermediate levels of acne severity (IG level 2). This may have reflected similar inaccuracies of clinical appraisals at this level or that intermediate grade image features provided insufficient feature variations for good distinction. The original training dataset may also have been underpowered at this level.
The authors noted that another limitation to their study was that acne severity was predicted inaccurately, which they said may reflect underrepresentation of darker skin types in the original training dataset. They concluded that larger patient enrollments for future validation studies could be expected to improve the performance of the algorithm-based automated grading system for acne vulgaris.
REFERENCES
Lim ZV, Akram F, Ngo CP, et al. Automated grading of acne vulgaris by deep learning with convolutional neural networks [published online September 29, 2019]. Skin Res Technol. doi:10.1111/srt.12794
