A novel multi-layer perceptron model for assessing the diagnostic value of non-invasive imaging instruments for rosacea

Introduction

Rosacea is a chronic inflammatory skin disease characterized by various signs and symptoms, including facial flushing, erythema, papules, pustules, and phyma (Tan et al., 2017a). Some of these features, such as erythema and papules, can easily be confused with other facial dermatoses, such as lupus, eczema, and acne. The correct diagnosis of rosacea depends largely on the clinician’s subjective perception and experience (Tan et al., 2017b; Thiboutot et al., 2020). In recent years, the convolutional neural network has been utilized to objectively assess and classify rosacea based on the clinical photos of rosacea patients (Binol et al., 2020; Zhao et al., 2021). These networks are established based on a single type of image and require a huge amount of photos for training the network. In clinical setting, a variety of skin imaging instruments, such as reflectance confocal microscopy (RCM), dermoscopy, and VISIA have been utilized to assist in the diagnosis and measurement of rosacea (Logger et al., 2019). Dermoscopy is the most widely studied imaging instrument with the advantage of clearly displaying skin structures, particularly blood vessels for rosacea, from the surface to the mid-dermis (Logger et al., 2019; Micali et al., 2018). The most characteristic imaging feature of rosacea under dermoscopy is the presence of vessels arranged in a polygonal network (vascular polygons), which corresponds to superficial telangiectasis surrounding the follicles in histopathology (Lallas et al., 2014; Sgouros et al., 2018). Rosette signs and Demodex tails may also be observed under dermoscopy but are not specific for rosacea (Liebman et al., 2011; Rubegni et al., 2013; Segal et al., 2010). VISIA is another commercially available high-resolution facial imaging system that is particularly useful for showing the deep vascular component presenting as background erythema (Wang et al., 2018). This technique can also clearly display telangiectasis and has been applied for rosacea diagnosis, severity assessment, and therapy monitoring in previous studies (Micali et al., 2018; Schoelermann et al., 2016). RCM provides real-time microscopic images of the different skin layers deep in the papillary dermis (González et al., 2003). Existing studies describing the use of RCM in rosacea have mainly focused on the Demodex inhabitation of sebaceous follicles (Ruini et al., 2017; Sattler et al., 2015). However, Demodex inhabitation can also be observed in many other skin diseases (such as acne), and the clinical significance of Demodex in rosacea remains under debate (Falay Gur et al., 2018). Taken together, although these three non-invasive skin imaging tools have been used for the diagnosis of rosacea, comprehensive studies evaluating the diagnostic value of these techniques are lacking.

A multilayer perceptron (MLP) is a type of feedforward artificial neural network, which was generally used for complex issues. It consists of input, hidden and output layers. The MLP can discover complex nonlinear relationships between factors (input) and outcomes (output) (Plumb et al., 2005). In clinical applications, the MLP model has irreplaceable advantages in the analysis of multiple data sources. In clinics, collecting comprehensive clinical data and images for all patients are quite difficult, and traditional analysis might fill in the missing data, which often introduces uncertainty and makes the prediction unstable. By contrast, the MLP does not require to fill in the missing data because these missing parts would be automatically replaced by other information that is more reliable and has greater potential for clinical application. As a result, in the current work we built an MLP model to assess the reliability of these three types of non-invasive imaging tools for the diagnosis of rosacea.

Materials & Methods

Results

Data analysis

Sampling

Except for the random deletion of variables, the data contained block deletion are shown in Fig. 1. Among all the 520 patients, 474 underwent dermoscopic examination; 374, RCM data; 434, VISIA analysis, and 511, clinical inspection. A total of 291 patients underwent all these four examinations.

Demographic data analysis

As shown in Table 1, no statistically significant difference was found in age between the rosacea and control groups. However, a statistically significant difference was noted in the sex distribution (p = 0.00178).

Table 1:

Demographic characteristics of patients and main imageological features of these three types of instruments in the rosacea and control groups.

Demographic characteristics	Rosacea	Control	P-value
Age	32.66 ± 11.54	31.5 ± 12.44	0.3247
Gender
Female	333	119	0.00178
Male	37	31

Imageological features		Rosacea		Control		P-value
		N	%	N	%
Dermoscopy
Background	Yellow	125	33.8	51	49.5	0.003
	White	98	26.5	47	45.6	<0.001
	Reddish	130	47.8	19	33.9	0.058
Blood vessels	Uniform distribution	151	56.1	16	25	<0.001
	Dotted	27	10	17	25.8	0.001
	Curved	5	1.9	3	4.5	0.2
	Linear	75	27.9	25	37.9	0.112
	Branched	175	47.3	27	26	<0.001
	Reticular	178	48.1	29	27.9	<0.001
Follicular findings	Yellowish halo	117	41.8	21	26.9	0.017
	Dilation of orifice	17	6.1	4	5.1	0.754
	Pustules	46	16.4	10	12.8	0.438
VISIA
Background erythema	Forehead	158	52.1	54	45.8	0.239
	Intercilium	235	77.6	68	57.6	<0.001
	Superciliary arch	228	75.2	58	49.2	<0.001
	Cheek	297	98	100	84.7	<0.001
	Perioral	142	46.9	47	39.8	0.192
	Chin	59	19.5	17	14.4	0.225
	Periorbital	265	87.5	85	72	<0.001
	Nose	240	79.2	79	66.9	0.008
	Preauricular	13	4.3	8	6.9	0.281
Overall erythema (Level 0–3)		2.426 ± 0.759		1.846 ± 1.039		<0.001
Papules and pustules (Level 0–3)		2.683 ± 1.190		2.214 ± 1.089		<0.001
Visible capillary		212	67.5	51	42.5	<0.001
RCM
Stratum corneum abnormality		57	20.9	40	39.6	<0.001
Spongiform edema		220	80.6	86	85.1	0.48
Thickness of epidermis (um)		69.975 ± 15.555		67.413 ± 14.623		0.354
Interface change		22	8.1	11	11	0.381
>3 Demodex mites in single follicle (superficial)		56	20.8	19	19.4	0.764
>3 Demodex mites in single follicle (deep)		62	23.8	14	15.1	0.077
Density of vascular (Level 0–3)		1.425 ± 1.169		1.145 ± 1.216		0.084
Vasodilation (Level 0–3)		0.822 ± 1.087		0.678 ± 1.13		0.255
Blood flow speed (Level 0–4)		3.009 ± 1.073		2.767 ± 1.06		0.031
Vascular morphology	Branched	75	30	17	18.3	0.029
	Linear	216	86.4	79	84.9	0.73
	Curved	10	4	6	6.5	0.338
	Reticular	34	13.6	4	4.3	0.015

DOI: 10.7717/peerj.13917/table-1

Imageological features of three types of instruments

The main imageological features of rosacea were identified and are listed in Table 1.

Dermoscopy.

Evenly distributed reticular blood vessels (vascular polygons) and a yellowish halo around the follicle were the most important imageological characteristics of rosacea under dermoscopy. The rosacea group exhibited significantly more branched and reticular blood vessels and fewer dotted capillaries than did the control group. A yellow and white background on the dermoscope can be helpful for excluding rosacea. Most cases of dermoscopy also exhibited pustules, which, however, did not show any significant difference between the rosacea and control groups.

RCM.

RCM examination of the rosacea revealed non-specific features. An important characteristic feature was the presence of capillaries and venules located in the dermis. However, neither the vascular density nor the vasodilation level was significantly different between the rosacea and control groups. The blood flow speed of patients with rosacea was significantly higher than that of controls, which might contribute to the background erythema in rosacea. Although likely incidental, reticular blood vessels (vascular polygons), as described on dermoscopy, were also observed under RCM (Video S1). Most RCM cases also exhibited telltale presence of Demodex mites within the follicular infundibulum. The percentage of hair follicles affected by three or more Demodex mites was higher in the rosacea group than in the control group; however this difference was not statistically significant.

VISIA.

Under the VISIA skin analysis system, rosacea was characterized by confluent and diffuse background erythema, telangiectasis, or a combination thereof. According to our results, telangiectasis was preferable to be located at the upper eyelid in rosacea, and the background erythema was more likely to implicate the superciliary arch, glabellum, periorbita, and nose. In addition, papulopustular lesions were detected using VISIA (Fig. 2).

Process and results of multi-layer perception analysis

As shown in Table 2, the AUROC of demographic data, clinical inspection, RCM, VISIA, and dermoscopy were 0.5159 ± 0.0597, 0.6901 ± 0.0903, 0.5259 ± 0.0597, 0.4736 ± 0.0648, and 0.7573 ± 0.0505, respectively. Among these, dermoscopy achieved the highest AUROC. The integration of all the five different data sources achieved a prediction accuracy of 0.8385 ± 0.0436, which was much higher than each of them alone. The same results were also observed in the Kappa coefficient; a single source of demographic data, clinical inspection, RCM, and VISIA, which showed agreement between the true label and model prediction class by chance (Kappa coefficient = 0). The highest Kappa coefficients were 0.9258 ± 0.0536 in the training set and 0.5478 ± 0.0851 in the validation set, achieved by the integration of all five different data sources (Table 3).

Discussion

The correct diagnosis of rosacea can sometimes be challenging in clinical settings. The most widely accepted diagnostic criteria was released in 2017 by the National Rosacea Society Expert Committee based on the clinical features of rosacea (Gallo et al., 2018). However, the accuracy of rosacea prediction based on clinical inspection alone was only 0.690 according to our MLP model. As a result, objective detection instruments are required to improve the diagnostic accuracy. Although few studies have focused on utilizing non-invasive imaging technologies for rosacea, they were designed mainly for research purposes (e.g., follow-up of new treatments) rather than for clinical diagnosis (Bageorgou et al., 2019; Schoelermann et al., 2016; Sparavigna, Tenconi & De Ponti, 2014). Additionally, most methods can only measure one or a few parameters of the complex symptoms of rosacea. Therefore, a combination of several instruments is desirable. In this study, we developed a novel MLP model that could comprehensively assess and integrate the characteristics of RCM, VISIA, and dermoscopy, and further screened possible candidate variables of each instrument for the diagnosis of rosacea.

Many previous reports have described the morphology of rosacea using RCM, dermoscopy, and VISIA (Lallas et al., 2014; Liebman et al., 2011; Turgut Erdemir et al., 2014; Turgut Erdemir et al., 2017). However, the sample sizes were relatively small and the inter-study outcome variability was large. A possible explanation for these discrepancies is that the measurement sites were not consistent. Based on our results, RCM examination of rosacea revealed mostly non-specific features, except that the abnormal stratum corneum represented an exclusionary feature for rosacea, which might be because rosacea barely affects the stratum corneum. One advantage of RCM is its ability to assess the blood vessels live. Increased cutaneous blood flow was observed in rosacea by RCM, but neither the density nor the diameter of the blood vessels revealed any significant statistical difference. The characteristic reticular blood vessels that were often observed under dermoscopy could also be notably seen in RCM, which has not been previously reported (Video S1). According to our results, dermoscopy and RCM findings were in concordance with each other in many cases, which could explain why the integration of RCM into MLP the model did not significantly improve the AUROC of prediction. Another advantage of RCM is its ability to detect the Demodex (Harmelin et al., 2014). Many previous studies have focused on quantifying Demodex in each follicle and the percentage of follicles affected by Demodex (Casas et al., 2012; Turgut Erdemir et al., 2017). However, Demodex inhabitation is not specific to rosacea and the clinical significance of Demodex colonization in rosacea is still under debate (Falay Gur et al., 2018; Lacey et al., 2018). Consistent with previous studies, most rosacea cases in this study also exhibited the presence of Demodex within the follicular infundibulum, which showed no statistical difference compared with that of the control group.

According to previous studies, the reticular blood vessels (vascular polygons) are the most significant dermoscopic characteristics of rosacea (Lallas et al., 2014). An interesting novel finding under dermoscopy was that most rosacea cases exhibited pustules that might be invisible to the naked eye. Further studies with larger sample sizes are encouraged to confirm these findings.

Because deep localized blood vessels could also be involved in rosacea, the RCM and dermoscopy are not always effective. As a result, some previous studies have attempted to utilize VISIA for the analysis of the overall erythema of rosacea (Micali et al., 2018; Schoelermann et al., 2016). The advantage of VISIA over other instruments, such as RCM and dermoscopy, is that VISIA can provide a full view of the whole face instead of focusing on a specific area. We used the VISIA system to evaluate the deep blood vessels of rosacea by assessing the confluent and diffuse background erythema, and found that the severity of the background erythema and cheek involvement were supportive evidence for rosacea diagnosis. Telangiectasis is another important presentation of rosacea under VISIA, which can often be observed in control subjects with sun-damaged facial skin (such as photoaging) or glucocorticosteroid-induced dermatitis (Helfrich et al., 2015; Lahiri & Coondoo, 2016; Wang et al., 2019). However, the telangiectasis displayed by VISIA for rosacea were preferably located in the upper eyelid (Fig. 2). Moreover, based on our results, the blood vessels on the superciliary arch are an important feature of rosacea in VISIA.

Similar to other well-known artificial intelligence studies, the exact factors that play their roles in MLP remain elusive. How to open this “black box” was still one of the biggest challenges in the field of artificial intelligence. In this study, we used LR regression analysis, which partly explained the influence of factors on the model. However, the potential interactions between these factors and their nonlinear effects require further exploration. Moreover, owing to the limited sample size and the different sources that may have duplicate information, our model has a trend of overfitting for the integrated multi-source model. Further studies with larger or multicenter databases are encouraged to validate and modify the MLP model.

Conclusion

RCM, dermoscopy, and VISIA can present several imageological features and were of certain value for rosacea diagnosis. The combined analysis of the three non-invasive imaging tools, namely dermoscopy, RCM, and VISIA, using our MLP model may be useful for improving the accuracy of diagnosing rosacea.

Supplemental Information