Measurement of psoriasis-affected area with artificial neural network

To the Editor: The Psoriasis Area and Severity Index (PASI) is widely used as the gold standard for psoriasis severity assessment. However, it is limited by its subjectivity and intra- and inter-rater inconsistency. One of the main reasons for this limitation is assessing the affected area of psoriasis. To overcome this, various artificial neural network (ANN)-aided PASI scoring systems have been studied. Recently, using a semantic segmentation network, comparable performance to human raters was demonstrated.1Breslavets M. Shear N.H. Lapa T. et al.Validation of artificial intelligence application in clinical dermatology.J Am Acad Dermatol. 2022; 86: 201-203Abstract Full Text Full Text PDF PubMed Scopus (2) Google Scholar We applied a state-of-the-art ANN that can reflect psoriasis characteristics and showed that it outperformed not only 3 board-certified dermatologists, but also all other ANN-aided PASI area scoring systems (Supplementary Table I, available via Mendeley at https://doi.org/10.17632/8nzj9y2rhm.4).1Breslavets M. Shear N.H. Lapa T. et al.Validation of artificial intelligence application in clinical dermatology.J Am Acad Dermatol. 2022; 86: 201-203Abstract Full Text Full Text PDF PubMed Scopus (2) Google Scholar, 2Schaap M.J. Cardozo N.J. Patel A. et al.Image-based automated psoriasis area severity index scoring by convolutional neural networks.J Eur Acad Dermatol Venerol. 2022; 36: 68-75Crossref PubMed Scopus (8) Google Scholar We created a dataset of 2232 plaque psoriasis images obtained from 149 patients. The dataset mainly consists of images of affected areas used for PASI scores, and facial images were excluded to maintain anonymity. The ground truth was obtained by directly annotating the psoriatic lesion in the images with agreement between 3 annotator dermatologists. To analyze the performance with our automated psoriasis detector (APD), these 3 dermatologists calculated the percentages of affected area using the palm method. Considering the severity via Physician's Global Assessment, the dataset was split into training and test sets consisting of 1731 and 297 images, respectively (see the flowchart in Supplementary Figs 1 and 2, available via Mendeley at https://doi.org/10.17632/8nzj9y2rhm.4). We applied the Cascade Mask R-convolutional neural network (CNN) with a Swin Transformer small backbone, which exhibited the best known performance regarding “Common Objects in Context” instance segmentation (Fig 1 and Supplementary Fig 3, available via Mendeley at https://doi.org/10.17632/8nzj9y2rhm.4).3Liu Z. Lin Y. Cao Y. et al.Swin transformer: hierarchical vision transformer using shifted windows.arXiv: Comput Vis Pattern Recognit. 2021; : 10012-10022https://doi.org/10.48550/arXiv.2103.14030Crossref Google Scholar With respect to ground truth, the PASI area score accuracy, mean absolute error (MAE) in percentage, and intraclass correlation coefficient (ICC) of APD were 95.96%, 1.23% ± 2.34%, and 0.973 (95% confidence interval [CI] 0.966-0.979), respectively, while those of the mean of the 3 dermatologists were 76.43%, 4.86% ± 5.59%, and 0.894 (CI, 0.558-0.957). The ICC between the 3 dermatologists was 0.781 (CI, 0.726-0.826) (Table I).Table IPerformance of automated psoriasis detector and dermatologistsPASI area score accuracy (%)MAE (SD)‡MAE and ICC of affected area were calculated as a percentage.ICC†ICC calculated with respect to ground truth.Affected area (95% CI)‡MAE and ICC of affected area were calculated as a percentage.PASI area score (95% CI)P valueAPD95.961.23 (2.34)0.973 (0.966-0.979)0.948 (0.935-0.958)<.001Dermatologist A∗Dermatologists calculated the percentage of affected area using the palm method.82.153.80 (5.36)0.795 (0.744-0.837)0.665 (0.592-0.727)<.001Dermatologist B∗Dermatologists calculated the percentage of affected area using the palm method.72.395.58 (7.69)0.801 (0.540-0.895)0.726 (0.521-0.830)<.001Dermatologist C∗Dermatologists calculated the percentage of affected area using the palm method.74.755.20 (6.24)0.844 (0.523-0.928)0.740 (0.561-0.834)<.001Mean of all dermatologists76.434.86 (5.59)0.894 (0.558-0.957)0.811 (0.615-0.892)<.001Bold = The level of significance was a P value less than .05.∗ Dermatologists calculated the percentage of affected area using the palm method.† ICC calculated with respect to ground truth.‡ MAE and ICC of affected area were calculated as a percentage. Open table in a new tab Bold = The level of significance was a P value less than .05. APD outperformed the dermatologists in PASI area score accuracy, MAE, and ICC. Compared to ground truth, all 3 dermatologists showed good reliability (0.75< ICC <0.90), while APD showed excellent reliability (ICC >0.90).4Portney L.G. Watkins M.P. 3rd ed. Foundations of clinical research: applications to practice. Vol. 892. Pearson/Prentice Hall, 2009Google Scholar The standard deviation of MAE also shows that APD can assess affected areas with consistency. The ANN model used in this study was Swin Transformer, which features long-range dependency, meaning that distant pixels in an image can be trained more effectively than in CNN.5Dosovitskiy A. Beyer L. Kolesnikov A. et al.An image is worth 16x16 words: transformers for image recognition at scale.arXiv. 2016; https://doi.org/10.48550/arXiv.2010.11929Crossref Google Scholar As psoriasis regions tend to be scattered, Swin Transformer allows outperformance of other methods (eg, color-based thresholding, CNN).1Breslavets M. Shear N.H. Lapa T. et al.Validation of artificial intelligence application in clinical dermatology.J Am Acad Dermatol. 2022; 86: 201-203Abstract Full Text Full Text PDF PubMed Scopus (2) Google Scholar, 2Schaap M.J. Cardozo N.J. Patel A. et al.Image-based automated psoriasis area severity index scoring by convolutional neural networks.J Eur Acad Dermatol Venerol. 2022; 36: 68-75Crossref PubMed Scopus (8) Google Scholar Our limitations were single-centered data acquisition of mainly Fitzpatrick skin types III and IV, and investigation of only plaque psoriasis. In addition, a patient photograph is required for APD analysis, which can increase the time and cost of the visit. Given these limitations, APD may initially find a role in clinical studies where precise measurement of psoriasis severity is the most important. However, with further advances, it may someday be used routinely in the clinic. None disclosed.