Automated grading of acne vulgaris by deep learning with convolutional neural networks

• DOI: 10.1111/srt.12794
• Publication Date: 2019-10-01T02:39:17Z
• Authors: Ziying Vanessa Lim; Farhan Akram; Cuong Phuc Ngo; Amadeus Aristo Wi...; Wei Qing Lee; Kaicheng Liang; Hazel Hweeboon Oon; Steven Tien Guan ...; Hwee Kuan Lee
• Keywords: Grading (engineering); Confusion; Repeatability; Ground truth; Raw score

Abstract Background The visual assessment and severity grading of acne vulgaris by physicians can be subjective, resulting in inter‐ intra‐observer variability. Objective To develop validate an algorithm for the automated calculation Investigator's Global Assessment (IGA) scale, to standardize outcome measurements. Materials Methods A total 472 photographs (retrieved 01/01/2004‐04/08/2017) frontal view from 416 patients were used training testing. Photographs labeled according IGA scale three groups clear/almost clear (0‐1), mild (2), moderate severe (3‐4). classification model a convolutional neural network, models separately trained on image sizes. then subjected analysis algorithm, generated scores compared clinical scoring. prediction accuracy each grade label agreement (Pearson correlation) two computed. Results best was 67%. Pearson correlation between machine‐predicted score human labels (clinical scoring researcher scoring) various input sizes 0.77. Correlation predictions with highest when using Inception v4 largest size 1200 × 1600. Two sets showed high 0.77, verifying repeatability ground truth labels. Confusion matrices show that performed sub‐optimally 2 label. Conclusion Deep learning techniques harnessing high‐resolution images large datasets will continue improve, demonstrating growing potential grading.