Table 1.
Summary of 33 Studies Reviewed on ML Applications to Psoriasis Evaluation and Management.
| Application category | Study | Objectives and results | Methods | Data set | Dataset source and annotations |
|---|---|---|---|---|---|
| A. Evaluation using skin images | |||||
| Identification and differential diagnosis of psoriasis lesions | Aggarwal (2019) | Demonstrated that data augmentation can improve machine learning image recognition of 5 dermatological diseases: acne, atopic dermatitis, impetigo, psoriasis, and rosacea. Each of the 5 diseases had an increase in AUC after data augmentation, with an average increase in AUC of 0.132 and SD of 0.033. The AUC with data augmentation was 0.87 for psoriasis | DCNNs with data augmentation | 332, 92, 138, 280, and 96 skin images for acne, atopic dermatitis, impetigo, psoriasis, and rosacea, respectively | Open-source dermatological images captured through DermNet, Dermatology Atlas, Hellenic Dermatological Atlas, and Google Images |
| Kim et al (2019) | Used smartphone-based multispectral imaging to discriminate between seborrheic dermatitis and psoriasis on the scalp. Demonstrated that machine learning-based methods for classification yielded better outcomes than conventional spectral classification methods, achieving a sensitivity of 65% to 75% and specificity of 70% to 80% | SVM, logistic regression (Logi), and multilayer perceptron (MLP) | 18 000 spectral signatures obtained from seborrheic dermatitis and psoriasis lesions on the scalp for training and 2000 spectral signatures for testing; followed by model validation with images from 60 patients in a clinical trial | Images captured via smartphone-based multispectral imaging system with an external CMOS camera; annotations confirmed by 3 different medical doctors | |
| Verma et al (2019) | Developed an ensemble data mining and machine learning method to classify 6 skin diseases: psoriasis, seborrheic dermatitis, lichen planus, pityriasis rosea, chronic dermatitis, and pityriasis rubra. The multi-model ensemble method achieved an accuracy of 98.64% | Classification and regression trees (CART), SVM, DT, RF, gradient boosted decision trees (GBDT) | 366 psoriasis histopathology images | Images from UC Irvine machine learning repository (Guvenir et al, 1998) | |
| Zhao et al (2019) | Developed a CNN to classify skin images of 9 common skin disorders (lichen planus, lupus erythematosus, basal cell carcinoma, squamous cell carcinoma, eczema, pemphigus, psoriasis, and seborrheic keratosis) as psoriasis vs. non-psoriasis. CNN classifier showed superior performance (missed diagnosis rate: 0.03, misdiagnosis rate: 0.04) than 25 Chinese dermatologists (missed diagnosis rate: 0.19, misdiagnosis rate: 0.10) in diagnosis of psoriasis on 100 clinical images | DCNN | 8021 skin images of 9 common disorders including 900 psoriasis images | Images collected by dermatologists at Xiangya hospital; Annotated by 3 dermatologists with > 10 years experience at Xiangya Hospital according to the corresponding medical record and pathology results | |
| Shrivastava et al (2016) | Developed a computer-aided detection system to classify skin images from psoriasis patients as healthy versus diseased using higher order spectra, texture, and color features, with a classification accuracy of more than 99% | Principal component analysis (PCA), SVM | 540 skin images (270 healthy and 270 diseased) from 30 psoriasis patients of Indian ethnic origin | Images captured and annotated by a dermatologist at the Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India | |
| Shrivastava et al (2016) | Assessed the reliability of a method developed in a previous study to classify skin images from psoriasis patients as healthy versus diseased, with a mean reliability index of 98.71% for 11 distinct data sizes | PCA, SVM | 540 skin images (270 healthy and 270 diseased) from 30 psoriasis patients of Indian ethnic origin | Images captured and annotated by a dermatologist at the Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India | |
| Shrivastava et al (2015) | Developed a computer-aided detection system to classify skin images from psoriasis patients as healthy vs. diseased using grayscale, color, redness, and chaoticness features, with a classification accuracy of more than 99% | PCA, SVM | 540 skin images (270 healthy and 270 diseased) from 30 psoriasis patients of Indian ethnic origin | Images captured and annotated by a dermatologist at the Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India | |
| Mashaly et al (2011) | Evaluated 4 techniques to classify skin images of 6 common papulosquamous skin diseases: psoriasis, lichen planus, atopic dermatitis, seborrheic dermatitis, pityriasis rosea, and pityriasis rubra pilaris. The rough sets method recorded the highest accuracy (78% to 94%) and sensitivity (825 to 96%) of segmentation compared with other techniques | Rough sets, Topological derivative, K-means clustering, watershed | 50 skin images from each of psoriasis, lichen planus, atopic dermatitis, seborrheic dermatitis, pityriasis rosea, and pityriasis rubra pilaris | Images from a dermatologic online image atlas, a dermatologic image database from the University of Iowa College of Medicine, and a dermatologic image atlas | |
| Lesion segmentation | Dash et al (2019) | Developed an automated psoriasis lesion segmentation method based on a modified U-Net architecture, referred as PsLSNet, achieving an accuracy of 94.8% with 89.6% sensitivity and 97.6% specificity | DCNN with a modified U-Net architecture (PsLSNet) | 5241 skin images of psoriasis lesions from 1026 psoriasis patients | Images captured and annotated by a dermatologist at the Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India |
| Pal et al (2018) | Developed an automated method to segment psoriasis skin biopsy images into dermis, epidermis and non-tissue regions, with the fully CNN method achieving an accuracy (defined as Ratio of Correct Pixel Classification) of 88% | DCNN | 90 psoriasis skin biopsy images | Psoriasis-affected tissue biopsies collected and clinically confirmed by a dermatologist of West Bengal, India, prior to imaging | |
| George et al (2017) | Developed an automated method for superpixel segmentation of psoriasis lesions in skin images, with a pixel accuracy of 86.99% | Multiscale superpixel clustering, K-means clustering | 676 psoriasis skin images from 44 psoriasis patients | Images obtained by clinical photographers at the Royal Melbourne Hospital, Australia, over 3 years; Automatically annotated using pixel-based skin segmentation method (George et al 2016) | |
| Jarad et al (2017) | Developed an automated method for psoriasis lesion segmentation in skin images using color spacing algorithms, with an accuracy of 95% | K-means clustering based on CIE Lab L*a*b color spaces | 80 psoriasis skin images (48 abnormal, 32 normal) | Images from database of psoriasis section of Ramadi Teaching Hospital, Ramadi, Anbar | |
| Shrivastava et al (2017) | Developed an automated method for segmentation of psoriasis lesions in skin images for accurate risk assessment using a Bayesian model, with a classification accuracy of 99.84% | SVM, DT, neural network | 670 cropped images from 110 psoriasis patient images (218 healthy, 29 mild, 138 moderate, 164 severe, 121 very severe) | Images captured and annotated by a dermatologist at the Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India | |
| Lu et al (2013) | Developed an automated method for segmentation of scaling in psoriasis skin images to evaluate disease severity | SVM | 103 psoriasis skin images | Images from University of Melbourne and St. Vincenťs Hospital Melbourne Department of Dermatology | |
| Taur et al (2006) | Developed an automated method for segmentation of psoriasis lesions in skin images using a MSSC | MSSC | 12 psoriasis skin images | Images from Taichung Veterans General Hospital, Taichung, Taiwan | |
| Taur (2003) | Developed an automated method for segmentation of psoriasis lesions in skin images into normal and affected regions | Neuro-fuzzy classifier | 3 psoriasis skin images | Images from Taichung Veterans General Hospital, Taichung, Taiwan | |
| Lesion severity and area scoring | George et al (2019) | Developed an automated method to score scale severity in psoriasis skin images using local descriptors, yielding a scale severity scoring accuracy of 80.81% | Bag-of-visual-words (BoVWs), SVM, RF | 96 psoriasis skin images | Images obtained by clinical photographers at the Royal Melbourne Hospital (RMH), Australia, over four years; Annotated by 3 dermatologists of RMH (correlation of 0.47, 0.51, and 0.34 between each dermatologist-pair) |
| Meienberger et al (2019) | Developed an automated algorithm to score psoriasis affected area, achieving an accuracy of more than 90% in 77% of the images and differed on average 5.9% from manually marked areas. The algorithm area estimates differed from physcians’ area estimates by 8.1% on average | DCNNs | 259 plaque psoriasis skin images from Caucasian patients (203 for training/validation, 56 for testing) | Frontal or dorsal photographs taken of Caucasian patients 18 to 80 years old with plaque psoriasis from University Hospital of Zurich Department of Dermatology in Switzerland; PASI assessed by physicians with more than 3 years of experience, supervised by a senior dermatologist | |
| Fink et al (2018) | Designed a total body imaging system that can also automate PASI measurements | Total body imaging system designed by authors | 10 psoriasis patients with 16 single body images per patient, captured using the authors’ total body imaging system | Total body imaging system created at University of Heidelberg Department of Dermatology, Germany; System’s PASI calculations compared to physicians’ assessments of 10 patients | |
| George et al (2018) | Developed a semi-supervised computer-aided system for automatic erythema severity scoring in psoriasis skin images, with a FI score of 0.71 for the random forest classifier | Patch-based dictionary learning, random forest, SVM, boosting | 676 psoriasis skin images from 44 psoriasis patients | Images obtained by clinical photographers at the Royal Melbourne Hospital, Australia, over 3 years | |
| Raina et al (2016) | Developed an automated method to score erythema severity in psoriasis plaque images, with good agreement with subjective assessment of erythema severity (kappa = 0.4203) | Linear discriminant analysis classifier | 80 psoriasis skin images from 20 psoriasis patients | Images taken from patients of Seton clinics including the University Medical Center Brackenridge Dermatology Clinic, Seton Family of Doctors at Hays, and Trinity Clinic; Labeled according to median rating by each of five dermatologists, including 1 attending and 4 experienced dermatology residents (intraclass correlation coefficient = 0.7306) | |
| Shrivastava et al. (2016) | Developed a psoriasis risk assessment system to risk stratify psoriasis severity from skin images | SVM, DT | 848 skin images (383 healthy, 47 mild, 245 moderate, 145 severe, 28 very severe) from Indian psoriasis patients | Images captured and annotated by a dermatologist at the Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India | |
| Shrivastava et al (2015) | Developed an automated method to risk stratify psoriasis severity from healthy and diseased skin images using color feature patterns, with an accuracy of ~99% | SVM, PCA | 540 skin images (270 healthy and 270 diseased) from 30 Indian psoriasis patients | Images captured and annotated by a dermatologist at the Psoriasis Clinic and Research Centre, Psoriatreat, Pune, Maharashtra, India | |
| Shrivastava et al (2015) | Reviewed technologies for psoriasis risk stratification in current and existing literature | Computer-aided diagnosis | N/A | N/A | |
| Fadzil et al (2013) | Developed an assessment method that incorporates 3D surface roughness with standard clustering techniques to objectively determine the PASI scaliness score for psoriasis lesions, with an accuracy of 94.12% | 3D surface roughness measurement, K-means clustering, fuzzy c-means clustering | 1999 psoriasis lesion images from 204 Malaysian psoriasis patients (1351 training, 648 test) | Images taken and annotated by dermatologists at Hospital Kuala Lumpar, Malaysia | |
| Savolainen et al (1997) | Developed a color segmentation method to assess involved surface area in psoriasis patients and compared performance to human eye assessments | CIA | 26 psoriasis patients with chronic plaque psoriasis (14 male, 7 female) | Images taken at Department of Dermatology and Venereology, Oulu University Hospital, Finland; Percent area of psoriasis involvement assessed simultaneously from projected slides of photographs by 2 dermatologists, 2 residents of dermatology, 2 nurses, and 2 medical students; Other PASI parameters assessed by a dermatologist | |
| Savolainen et al (1998) | Compared surface area estimates by human eye versus computer image analysis using color segmentation. Human eye estimates were higher than computer estimates, tending to overestimate in cases where the PASI was under 15 | CIA | 15 psoriasis patients (14 male, 1 female) | Images taken at Department of Dermatology and Venereology, Oulu University Hospital, Finland; PASI and skin area assessed by by dermatologist | |
| Gomez et al (2007) | Compared available change detection techniques in the visualization and quantification of bi-temporal psoriasis images | Simple image subtraction, PCA, post-classification comparison, MAD transform | 6 temporal series of psoriasis images (each consisting of 4 images collected in 1 week interval) | Images taken from Gentofte Hospital, Denmark; Lesion severity scored by dermatologists, and ground-truth change calculated by subtracting dermatologists’ scores at 2 different time instants | |
| B. Clinical management | |||||
| Prediction of complications | Munger et al (2019) | Used ML methods to determine top predictors of non-calcified coronary burden by Coronary Computed Tomography Angiography in psoriasis. These factors were related to obesity, dyslipidemia, and inflammation | RF | 263 consecutive patient records (January 2013-January 2018) with 62 available variables measured at baseline from the Psoriasis Atherosclerosis Cardiometabolic Initiative | Data from Psoriasis Atherosclerosis Cardiometabolic Initiative; Annotations obtained from patient records |
| Patrick et al (2018) | Used genetic data to assess the risk of psoriatic arthritis development in psoriasis patients. Identified 9 new loci for psoriasis or its subtypes and achieved 0.82 AUC in distinguishing PsA versus PsC when using 200 genetic markers | RF, conditional inference forest, shrinkage discriminant analysis, elastic net regression | Six cohorts with >7000 genotyped psoriatic arthritis and cutaneous-only psoriasis patients | 5 GWAS datasets: CASP, Exomechip with with GWAS content, Genizon, Kiel, PsA GWAS; 1 Immunochip dataset: PAGE | |
| Treatment | Patrick et al (2019) | Predicted drugs that can be repurposed to treat immune-mediated cutaneous diseases like psoriasis. The method confirmed drugs that are known to be effective for psoriasis and identified potential new drug candidates currently used to treat other diseases | NLP with word embeddings | 353 drugs indicated as not being used to treat psoriasis but used to treat some other immune-mediated disease, evaluated using a database of 20 million MEDLINE abstracts (3.3 billion words) | Database of 20 million MEDLINE abstracts (3.3 billion words) |
| Tomalin et al (2019) | Developed models to predict long-term treatment response to tofacitinib and etanercept though quantification of 157 inflammatory and cardiovascular proteins in the blood of psoriasis patients. Their models accurately predict the 12-week clinical endpoint for psoriasis following tofacitinib (auROC = 78%) or etanercept (auROC = 71%) treatment in a validation dataset | Prediction of Microarrays (PAM), Threshold Gradient Descent Regularization (TGDR), Generalized Linear Models (GLMnet), Partial Least Squares (PLS), Neural Networks (NNET), SVMs, RFs | 157 disease proteins measured from the blood of 266 patients with moderate-to-severe psoriasis | ClinicalTrials.gov: NCT01241591 (Bachelez et al, 2015) comparing efficacy and safety of tofacitinib and etanercept; PASI75 (outcome variable: patient labeled as a responder if PASI decreases by >75% after 12 weeks of treatment) determined by dermatologists | |
| Zhang et al (2014) | Used an NLP method to identify known and unknown drug-drug interactions from MEDLINE and EHR data | NLP with semantic predications | 224 unique drugs from 22 patients, evaluated using the SemMedDB database (21 million citations and 119 million sentences) | SemMedDB database (21 million citations and 119 million sentences); Salient predications validated by physicians | |
Abbreviations: AUC, area under the curve; CIA, computer image analysis; DT, decision tree; CNN, convolutional neural network; DCNN, deep convolutional neural network; MSSC, multiresolution-based signature subspace classifier; NLP, natural language processing; PCA, principal component analysis; RF, random forest; SVM, support vector machine.