Smartphone-based multispectral imaging and machine-learning based analysis for discrimination between seborrheic dermatitis and psoriasis on the scalp

For appropriate treatment, accurate discrimination between seborrheic dermatitis and psoriasis in a timely manner is crucial to avoid complications. However, when they occur on the scalp, differential diagnosis can be challenging using conventional dermascopes. Thus, we employed smartphone-based multispectral imaging and analysis to discriminate between them with high accuracy. A smartphone-based multispectral imaging system, suited for scalp disease diagnosis, was redesigned. We compared the outcomes obtained using machine learning-based and conventional spectral classification methods to achieve better discrimination. The results demonstrated that smartphone-based multispectral imaging and analysis has great potential for discriminating between these diseases. © 2019 Optical Society of America under the terms of the OSA Open Access Publishing Agreement


Introduction
Seborrheic dermatitis and psoriasis are common chronic inflammatory skin diseases that present with erythematous scaly skin lesions.Approximately 3% of individuals worldwide currently experience these diseases, and more than 3 million individuals are treated in the USA and United Kingdom each year.Seborrheic dermatitis typically occurs on the scalp or the face, usually manifesting as oily and greasy inflammation.Its symptoms include itching, reddish skin, and flaking on greasy skin areas [1].Conversely, psoriasis, which is now considered as a multifactorial immune-mediated disease, is manifested by itching, reddish skin, and silvery-white scaly plaques on the scalp, face, or extremities [2,3].
Although neither seborrheic dermatitis nor psoriasis is fatal to humans, both diseases are associated with significant physical and psychological burden.Accurate/preemptive diagnosis of these diseases is thereby crucial to control their symptoms and predict their prognosis, in terms of their chronic and relapsing properties and associated comorbidities, such as metabolic syndrome [4,5].For this purpose, an advanced tool capable of diagnosing the diseases at home, a local hospital, or elsewhere with high accuracy would be beneficial for better treatment and management.However, using a conventional dermascope, the discrimination between seborrheic dermatitis and psoriasis has been clinicopathologically very challenging.In particular, when these diseases occur only on the scalp, the differential diagnosis between them is more challenging even among experienced physicians owing to their similar features [6,7].
Recently, dermatologists have been exploring new methods for better diagnosis of seborrheic dermatitis and psoriasis.Among them, a videocapillaroscope, which provides high magnification, has been used to examine and classify the morphological shape of the capillaries to distinguish the diseases accurately [8,9].However, the shape of the capillaries is immensely sensitive to external environments, and as a result, the outcomes have been shown to vary [10].Moreover, the diagnosis using the device can be only realized by dermatologists in a hospital.Thus, it may not be suited for preemptive diagnosis of the diseases under ubiquitous environments.Therefore, development of more advanced tools is needed for better mobile diagnosis of seborrheic dermatitis and psoriasis on the scalp.
Spectral imaging and analysis techniques have been widely used for the non-invasive diagnosis of diseases occurring on the skin and scalp [11,12].In particular, they offer high sensitivity and specificity in the detection of various diseases, which is not possible using monochrome and RGB color [13][14][15][16].For this reason, spectral imaging and analysis may also be applicable in discriminating between seborrheic dermatitis and psoriasis on the scalp with high accuracy, since it can differentiate slight spectral differences between diseases with different biological compositions.However, conventional spectral imaging systems are relatively bulky; thus, they need to be miniaturized and combined with information technology (IT) for self-diagnosis of these diseases.Therefore, for the expedient selfdiagnosis of seborrheic dermatitis and psoriasis on the scalp, a new type of multispectral imaging system based on IT as well as a relevant spectral analysis technique is needed.
Thus, we built a smartphone-based multispectral imaging system for scalp care and investigated the potential of various multispectral imaging and analysis methods to discriminate between seborrheic dermatitis and psoriasis on the scalp using the system.The smartphone-based multispectral imaging system consisted of an external CMOS camera and a light-emitting diode (LED) array, including narrow bandpass filters.The system was connected to a smartphone for image acquisition, image transfer, and image analysis in a server and classified image display [17].After evaluating the performance of the system, it was employed to discriminate between seborrheic dermatitis and psoriasis on the scalp of patients.Moreover, to achieve a more accurate diagnosis of the diseases, various machine learning-based and conventional spectral imaging and analysis techniques were applied for spectral classification, and their outcomes were compared.

Methods
We built a smartphone-based multispectral imaging system with an external CMOS camera for the diagnosis of seborrheic dermatitis and psoriasis on the scalp.This system consisted of a multispectral imaging module, which included an LED array, various optical components, a CMOS camera (See3CAMCU50, E-CON SYSTEMS) connected to the smartphone, and an interface circuit for synchronization between the CMOS camera and the LED array, and a system control/image analysis module, which included a smartphone (SM-G935, SAMSUNG) and a server [Fig.1(a)].
The LED array included nine white LEDs (Iws-351-white, ITSWELL).Eight of the LEDs were utilized for multispectral imaging, and one LED was utilized to examine the regions of interest.After a linear variable filter (LF102499, DELTA OPTICS) was diced to construct narrow bandpass filters, each diced linear variable filter was placed prior to each LED to select the wavelengths of the light emitted from the LED.The sizes of the diced filters were measured to be 3.8 mm (width) × 3.3 mm (height) × 2.5 mm (thickness).The center wavelengths of the filters were ~453 nm (FWHM: 9.3 nm), ~493 nm (FWHM: 11 nm), ~520 nm (FWHM: 12.3 nm), ~545 nm (FWHM: 13.3 nm), ~580 nm (FWHM: 15.6 nm), ~605 nm (FWHM: 16.2 nm), ~638 nm (FWHM: 16.6 nm), and ~663 nm (FWHM: 16.8 nm) [Fig.1(b)].The filtered light emitted from the selected LED was illuminated onto the skin regions of interest on the scalp after passing through a polarizer oriented in the vertical direction.The reflected light from the skin passing through another polarizer, oriented in the horizontal direction, wa pixel color CM rejecting spec In the sys and the LED A Bluetooth l Macroblock) Instruments), received com turned on/off utilized to sup 5 V was supp V was downshifter adjust microprocesso interface circu In the syn controlled the controlled the the Bluetooth from the sma was selected, the smartphon a commend communicatio For multis i.e., an image selected by tu to a server thr previously.T models in the in the smartp and image pr the image of t

Spectral classification using conventional linear distance measures and machine learning techniques
To discriminate between seborrheic dermatitis and psoriasis on the scalp, machine learning techniques, including the support vector machine (SVM), logistic regression (Logi), and multilayer perceptron (MLP) techniques, and conventional linear distance measurement methods, such as the Euclidean distance (ED) and spectral angle mapper (SAM), were applied for spectral classification.Before the obtained image cube was applied for spectral classification, it was preprocessed via calibration, as shown previously [19].For spectral classification of the spectral signature of every pixel, the similarity between the spectral signatures and reference signatures, extracted from the skin lesions of interest, was calculated using the trained model based on the machine learning techniques and the distance measurement methods [20,21].
In the conventional linear distance measurement methods, the smaller the similarity values of each reference, the closer it is to the reference signatures.The similarity between the spectral signatures and reference signatures calculated using the ED and SAM can be demonstrated as the straight distance (ED) and the angle (SAM) between two vectors [Figs.

2(a) and (b)] [19]
. However, their construction model procedures are quite different from the machine learning-based techniques.For spectral classification using the machine learning techniques, a training data set, which includes the ground truths in advance, was first trained using the methods to construct models suited for the classification of two reference signatures with high accuracy, unlike the traditional methods with fixed models for calculating the similarities.The spectral signature was considered to be a vector, and the similarity was calculated using an equation derived from their own function in the machine learning techniques.In the machine learning techniques, the process for constructing each machine learning model was different from each other.For example, Logi uses the sigmoid function for building an optimized model for the binary classification [Fig.2(c)], whereas the SVM finds the best hyperplane, which can distinguish the given data set successfully with low errors [Fig.2(d)].Conversely, the node, known as neuron, in the MLP model receives some features from the previous nodes by multiplying its weights with the activation function, and then the final output of the transferred values is compared to the ground truths [Fig.2(e)] [22].The machine learning models were implemented into the server using Scikit-learn library.Here we determined the key parameters for each method to achieve the best outcome using a grid search method [23].For the logistic regression, the inverse of regularization strength parameter was determined to be 0.0303.For the support vector machine, the kernel type was linear, and the penalty parameter C was 0.1.For the multilayer perceptron, the hidden layer size was 4, 4, 4, the activation function was 'Identity', the learning rate was 1.0, and the optimizer type was 'Adam'.
For the training and testing of the machine learning models, 18000 spectral signatures obtained from the lesions of interest were utilized as the training set, while the unused 2000 spectral signatures were utilized as the test set.In addition, 10-fold cross-validation was performed to assess the generalization of the models [20].The cross-validation technique is a conventional validation technique for quantitatively evaluating the outcomes of a trained model.

Discrimin via multispec
Clinical trials     For the quantitative analysis of the models, the ROC-AUC, accuracy, sensitivity, and specificity of the models were compared.All the models showed great performance in the discrimination between the diseased and normal skin regions.Among them, the SAM model yielded the best outcomes in the discrimination between the diseased and normal skin regions, unlike in the discrimination between psoriasis and seborrheic dermatitis.The ROC-AUC, accuracy, sensitivity, and specificity of the SAM model were 0.98, 95, 90, and 100 for the discrimination between psoriasis and normal skin.Further, the ROC-AUC, accuracy, sensitivity, and specificity were 1.00, 100, 100, and 100 for the discrimination between seborrheic dermatitis and normal skin (Tables 2 and 3).The ROC-AUC and accuracy of the SAM model were higher than those of the other models.

Discussion
In this study, we demonstrated the potentials of mobile multispectral imaging and analysis for discriminating between seborrheic dermatitis and psoriasis on the scalp.The developed system was capable of obtaining multispectral images within the wavelength range of 453 -663 nm.The multispectral images were recorded using developed android applications.An image cube could be transferred to a server for discriminating between seborrheic dermatitis and psoriasis on the scalp, followed by the display of the resulting image on the smartphone.Moreover, it allowed self-monitoring of the diseased regions on the scalp through the smartphone.A previous report has shown that a smartphone-based multispectral imaging system has the potentials for diagnosing various skin lesions.However, the previous system showed some limitations in the examination of skin lesions [19].For example, when the system was applied to examine skin lesions on the scalp, patients could not monitor the skin lesions of interest on a smartphone display.To address this limitation, an external CMOS camera was incorporated to the developed system for the self-diagnosis of skin lesions on the scalp rather than utilizing the camera in the smartphone.This enabled diseased regions on the scalp to be monitored in the smartphone display.
When the performance of the developed system was evaluated, a frequency line set at 45 cycles per millimeter could be clearly resolved in the vertical and horizontal directions by the system.In our previous study, the smartphone-based system was able to resolve a frequency line set at 51 cycles per millimeter [19].The spatial resolution achieved by our proposed system was slightly lower than that by the previous system.It could be improved by utilizing a better external CMOS camera and a lens with a higher numerical aperture.Conversely, the system we developed herein also showed a slightly better performance than did the LCTFbased multispectral imaging system in multispectral imaging and analysis as shown in Fig. 4. The spectral classified images obtained using the LCTF showed more shot noise regions, more misclassified regions at the boundary of the letters, and lower contrast than did those obtained using our system.These might be because of the low signal-to-noise ratio in the detection of light owing to the low light transmission capability of the LCTF system.
In the discrimination between seborrheic dermatitis and psoriasis on the scalp using the system, the machine learning-based classification methods showed a relatively better performance than did the conventional spectral classification methods, particularly the SAM (Fig. 5).The spectral classified images of the psoriasis regions exhibited only a few misclassified regions, while most regions correctly indicated psoriasis.Therefore, via a simple quantification of the spectral classified images and appropriate threshold, such as calculating the ratio of the number of psoriasis and seborrheic dermatitis pixels, we could determine whether the skin lesions were psoriasis.Also, in the classified images for the seborrheic dermatitis [Fig.5(a), bottom], we could still distinguish and classify the diseased regions as seborrheic dermatitis.The images of the patients with seborrheic dermatitis and psoriasis were obtained at a quite early stage; thus, it was challenging to classify between them even by an experienced clinician.Thus, these results suggest that the smartphone-based multispectral imaging and analysis may have a potential use for discriminating between seborrheic dermatitis and psoriasis in mobile environments.
For spectral classification, the ED, SAM, SVM, Logi, and MLP classification models were adopted.The machine learning models have previously been successfully utilized for detecting various diseases, including ovarian cancer [24], colonic and gastric cancers [25], prostate cancer [26], and Alzheimer's disease [27].In previous studies, the machine learning models yielded strong classification outcomes for the detection of diseases.As in a previous study, three of the machine learning models constructed via multispectral imaging herein showed acceptable classification performance.The SVM model achieved the highest accuracy in the discrimination between them; it also had a better ROC-AUC (0.76) than the multilayered perceptron model (Table 1).In the comparisons between the three machine learning-based models and two conventional classification models, the sensitivities of the former were higher than those of the latter.The ROC-AUCs achieved by the machine learning models were also higher than those by the conventional classification models (Table 1).These results suggest that the machine learning-based models may provide better sensitivity and specificity compared with the conventional classification models; however, further clinical studies with larger numbers of patients are needed to confirm this.
In this study, to validate the performance of the constructed model under such conditions, 10-fold cross validation was performed owing to the relatively small number of patients with seborrheic dermatitis and psoriasis.This method has been typically utilized to evaluate the generalization and reproducibility of the machine learning-based models statistically.The specificity and sensitivity of the constructed models for discriminating between seborrheic dermatitis and psoriasis on the scalp were considerably high.Therefore, these results and consistent trends may provide confidence in the ability of multispectral imaging and analysis in discriminating between seborrheic dermatitis and psoriasis on the scalp based on the proposed models.

Conclusions
In conclusion, we employed multispectral imaging and analysis for discriminating between seborrheic dermatitis and psoriasis on the scalp using the smartphone-based multispectral imaging system we developed on the basis of various classification models, such as the ED, SAM, SVM, Logi, and MLP.The developed system was capable of performing multispectral imaging and analysis of lesions on the scalp for self-monitoring via a smartphone display.Herein, the machine learning methods (Logi, SVM, and MLP) for spectral classification showed a sensitivity of 65, 75, and 75 and specificity of 80, 70, and 70, respectively.The conventional methods (ED and SAM) showed a sensitivity of 70 and 70 and specificity of 70 and 75, respectively.Herein, the machine learning-based models yielded better outcomes than did the conventional methods, demonstrating the potential as a mobile detection device.Taken together, these results suggest that the multispectral imaging and analysis based on the machine learning techniques using the developed system would be a valuable tool for clinical diagnosis nearly in any setting; however, extensive clinical trials with a large number of patients are needed to confirm this.
Fig. derma imagi mm × smartp Fig. 2 mappe After the removed the h spectral class estimate the p

3 .
Analysis of a hi and (b) intensity p LCTF-based m tem including 1), 12 high-in ra (IDS, UI-33 em, the light em arizer in the ps and then pass was then record shows the spe rs "M," " on target obtain ntal dotted lin in the vertical m [Fig.3(b)].T meter.e of the resolution ne B' (right).CTF-based mul , an LCTF (T 1), and a mon CTF-based mul to a target afte get was collect r in the s-direc le within the r n red, green, b Fig. 4 tunabl magni mm x Fig. 5 distan suppo psoria multis magni s on the scalp.
Figure 6 illus normal skin images shown regions in mo and (b) exhib blue, and gree respectively.misclassified seborrheic de overall region the spectral cl was observed misclassified the models ha the normal sca

Table 2 . Quantitative analysis of the constructed model for psoriasis versus normal skin a
using the Euclidean distance (ED), spectral angle mapper (SAM), logistic regression (Logi), support vector machine (SVM), and multilayer perceptron (MLP) with reference spectral signatures in the discrimination between psoriasis and normal skin on the scalp (total number of patients = 40, number of patients with psoriasis = 20, and number of normal subjects = 20).ROC-AUC, area under the receiver-operating characteristic curve b