Validation of Artificial Intelligence Application in Clinical Dermatology
An ANN showcased enhanced reliability over human experts in evaluating skin affected by psoriasis
January 25, 2021
Denys BreslavetsMaksym BreslavetsNeil H. ShearTatiana LapaAlina Breslavets
ann in dermatologydermatological researchpsoriasis diagnosisai precisionskin condition evaluation
This Research Letter, titled 'Validation of Artificial Intelligence
Application in Clinical Dermatology,' was originally published in
the Journal of the American Academy of Dermatology (JAAD) in January
25, 2021. The following is a reproduction of the original article.
To the Editor: Artificial intelligence application may play an
essential role in clinical medicine and has previously been evaluated
to identify malignant skin lesions and assist with triage of
patients.1, 2 Meanwhile, the dermatologic examination is
still subjective and depends on visual assessment, previous training,
and experience.3 We conducted an independent retrospective
study to compare the performance of a board-certified dermatologist,
two university-level science students, and an artificial neural
network (ANN) in assessing skin areas affected by psoriasis through
calculating the percentage of the affected skin, which is routinely
measured by the palm method.4
The performance of participants was analyzed on 118 de-identified, non
altered color photos of psoriasis at a resolution of 3024 × 4032
pixels. These photos were selected from a database of the single,
community-based dermatology clinic. The participants who assessed the
photographs had to calculate the percentage of affected area. A data
set of an additional 2094 unaltered photos of psoriasis and
corresponding manually colored image masks was used to train the ANN.
We used the semantic segmentation approach because it segments out the
objects of interest (skin with the disease) and the rest of the image.
The trained ANN was able to calculate the psoriasis-affected
percentage of skin in any provided photograph (Figs 1 and 2).
Potentially, image segmentation techniques such as clustering can be
used. However, the presentation of psoriasis can be variable with a
spectrum of colors attributed to the disease that makes clustering
segmentation less suitable in the current context.
Fig 1: Pre-processing image analyzed by the artificial neural
network.
The ANN noticeably outperformed the human participants and had very
short execution time (within 1 second), predicting the affected
percentage accurately and precisely. The ANN had a mean percentage
error (MPE) of 8.71%, with a standard deviation (SD) of 6.70% (95%
confidence interval [CI], 7.64%-10.02%) compared with the physician's
MPE of 28.16% (SD, 22.69%; 95% CI, 24.76%-32.61%). We observed that
the ANN was more consistent at predicting the affected area than an
experienced dermatologist. Participants 1 and 2 showed an MPE of
30.13% (SD, 39.05%; 95% CI, 24.58%-39.30%) and an MPE of 159% (SD,
120.9%; 95% CI, 138.2-184.0), respectively. Moreover, the trained
model was able to determine well-defined skin lesions in such
conditions as mycosis fungoides and nummular dermatitis.
Fig 2: Post-processing image analyzed by the artificial
neural network.
Our study has limitations. First, certain locations and types of skin
lesions make it challenging to recognize them (scalp, genital areas,
ill-defined rashes). Second, assessment in different Fitzpatrick skin
types may require further adjustment. Third, a limited number of
psoriasis images was used. Finally, our study was conducted at a
single medical center.
The study results show that a trained ANN was able to calculate the
affected skin surface area consistently and precisely, outperforming a
dermatologist and untrained laypersons. Implementation of artificial
intelligence can be a valuable tool for fast and precise estimation of
skin areas affected by psoriasis. Also, the scalability of the ANN and
the ability to use it as a mobile application opens the horizon for
point-of-care assessment, including the analysis of the standard front
and back clinical images that would be sufficient to calculate the
whole-body surface area.5
References
1
Phillips, M., Greenhalgh, J., Marsden, H., & Palamaras, I. (2019). Detection of malignant melanoma using artificial intelligence: An observational study of diagnostic accuracy. Dermatology Practical & Conceptual. https://doi.org/10.5826/dpc.1001a11
2
Du‐Harpur, X., Watt, F. M., Luscombe, N. M., & Lynch, M. D. (2020). What is ai? applications of artificial intelligence to dermatology. British Journal of Dermatology, 183(3), 423–430. https://doi.org/10.1111/bjd.18880
3
Breslavets, A., Breslavets, M., & Shear, N. H. (2019). Quantification of randomness (entropy) as a clinical tool to assess the severity of skin disease. Medical Hypotheses, 132, 109311. https://doi.org/10.1016/j.mehy.2019.109311
4
Amirsheybani, H. reza, Crecelius, G. M., Timothy, N. H., Pfeiffer, M., Saggers, G. C., & Manders, E. K. (2001). The natural history of the growth of the hand: I. hand area as a percentage of body surface area. Plastic and Reconstructive Surgery, 107(3), 726–733. https://doi.org/10.1097/00006534-200103000-00012
5
Kreft, S., Kreft, M., Resman, A., Marko, P., & Kreft, K. Z. (2006). Computer-aided measurement of psoriatic lesion area in a multicenter clinical trial—comparison to physician’s estimations. Journal of Dermatological Science, 44(1), 21–27. https://doi.org/10.1016/j.jdermsci.2006.05.006
Authors
1
Maksym Breslavets, MD, PhD, FRCPC - Centre for Medical and Surgical Dermatology, Whitby, Ontario, Canada
2
Neil H. Shear, MD, FRCPC - Faculty of Medicine (Dermatology), University of Toronto, Toronto, Ontario, Canada
3
Tatiana Lapa, MD - Dalla Lana School of Public Health, University of Toronto, Toronto, Ontario, Canada
4
Denys Breslavets - Faculty of Science, Ryerson University, Toronto, Ontario, Canada
5
Alina Breslavets - Faculty of Science, University of Ottawa, Ottawa, Ontario, Canada
Acknowledgements
1
Author bios are presented as they were at the time of the article's original publication
