Artificial intelligence in the management of acne

Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical industry. In this episode we will talk about how artificial intelligence has contributed to the dermatologist's job by helping the assessment of active acne lesions and scars.

Did you know that acne is prevalent in 9.4 percent of the global population and it is identified as the 8th most common disease worldwide?3

Today we will present how artificial intelligence can be applied in dermatology, particularly in the detection and classification of acne vulgaris, usually referred to as acne.1

You may be familiar with the use of medical images in artificial intelligence, mostly MRI, X-ray, and CT for other specialties. But an important source of information for dermatology can come from photos taken by smartphones, or other portable cameras, because they are the most accessible and the most convenient way to monitor patients outside the hospital, especially in patients with skin diseases.1

This becomes very useful for acne patients. Let's remember that acne is one of the most prevalent skin conditions, which is typically seen among almost 90 percent of adolescents, but it can persist into adulthood in about 12 to 14 percent of the cases.1,3 It is widely known that acne harms not only the patients’ physiological conditions, but also their mental health, by laying detrimental effects on self-esteem.1

More importantly, the early diagnosis and accurate continuous self-monitoring of acne is key to control and alleviate the discomfort. In this case, self-monitoring includes a broad range of concepts, including self-diagnosis, self-evaluation of the effect of treatment or therapy provided by the dermatologists, etc. Hence, self-monitoring helps both the patients and the dermatologists.1

Therefore, to better monitor the disease, there are currently comprehensive, easy-to-use guidelines for estimating acne severity based on counting the number of lesions, either papules or pustules, per half face. Here is where smartphone images become very useful, since they can help the research using image analysis of different skin conditions. In fact, plenty of work using the power of artificial intelligence or convolutional neural networks has been done, and these methods reach an accuracy of up to 93 percent, and often even outperform dermatologists themselves.1

However, most of these studies have a common weakness, interpretability, a key characteristic that makes predictions more convincing. In this case, interpretability refers to the fact that these studies produce solely the number of acne lesions or the level of acne severity.1

Building up on these weaknesses is where the work of Hao Wen and colleagues becomes relevant. The researchers from the Department of Computer Science and Technology in the Tsinghua University of Beijing published in the Technology and Health Care journal a study focused on the application of artificial intelligence for the automatic location of acne lesions in facial images and then, the acne severity and lesion counting could be determined.1

To achieve this, the authors take advantage of object detection methods, also known as object localization methods. These procedures are the most straightforward and perhaps the most convincing for acne analysis on facial images. Once the acne lesions are located in the face, a numerical scoring could be applied to make assessments of acne severity.1

Object detection has achieved great success in many application areas, for instance in autonomous driving, but its potential in facial acne analysis seems to have not been fully exploited so much before. Probably because there are some drawbacks, like their time efficiency, which of course could be a serious issue for autonomous driving, but is not a big issue for acne scenarios.1

Hao Wen and colleagues used a dataset called ACNE04, which contains about 1 400 images that are each annotated with bounding boxes on each lesion, as well as the total number of lesions and the severity. After checking the characteristics of each image, they removed images with either low quality or not suitable enough for the analysis because the facial skin made up too small areas. This filtering left 1,222 high quality images for training and validation.1

These filtered images were used as input for the model developed by the authors, and after testing, they obtained average precisions of up to 0.536. To have more context, consider that the precision reached by other models was 0.526, in this case the authors used one called Yolov4.1

In addition, remember that the number of acne lesions is the essential step for estimating acne severity in some numerical criteria, the model thus also counted the predicted bounding boxes in the images. The counts were further used to compare the performance of the model using something called bias arrays, which is obtained by subtracting the predicted acne lesion numbers by the ground truth acne lesion numbers. With these values, the authors confirmed that among all the models tested, the one implemented by them gives the least difference in lesion counts compared with the ground truth.1

Afterwards, the authors predicted the acne severity values and compared them to the ground truth, showing that predictions for low severity level images are considerably accurate, while the only flaw lies in the predictions for the very severe group.1 Another limitation of this study is that they worked with only one class for acne, when it can be further classified into six or more subtypes, and even include acne scars.1

In summary, the model proposed by Hao Wen and colleagues could help patients and perhaps their dermatologists to conduct continuous monitoring outside of the hospital, evaluating the facial acne states, such as, acne lesion number, severity, as well as progression of the acne and estimation on therapy efficacy.1

CAPSULA

Acne is an almost universal condition in young people. Although information about its contributing factors is limited, it is of great importance for adolescents to be conscious about the factors that aggravate or ameliorate this condition, in order to seek medical help on time, to prevent scarring and costly treatments.2

With this in mind, a survey on 500 secondary school students, the majority being 15 years old, was used to gain a better understanding of adolescents’ beliefs about acne. The survey showed that almost 80 percent of those with acne did not seek medical help.2

Likewise, 85 percent of the students believed that inadequate face washing contributes to acne, followed by hormones with 84 percent and consuming sweets with 82 percent. Also, more than two thirds of them believed that consumption of greasy food, makeup, stress, and sweating aggravate acne. In contrast, most believed in the benefits of cosmetic treatment of acne. They also believed that an increase in water consumption, a change to a healthy diet, and being on school holidays can ameliorate the condition.2

This shows that knowledge about the factors that aggravate and ameliorate acne is insufficient among students, regardless of the presence of acne, and it is the reflection of misconceptions. Therefore, more efforts are needed to educate adolescents about acne.2

END OF CAPSULA

Welcome again. In the previous section, we talked about the application of artificial intelligence on the detection of acne lesions and the prediction of acne severity. Now we will talk about how artificial intelligence can also be useful to classify acne scars.

Recapitulating a little, we have talked about the importance of detecting acne lesions in time for proper treatments. However, the scars that originate from the damage to the tissue, caused by acne need long-time treatment to be removed. In fact, acne scars can be more troublesome for some patients than the acne itself, and they may have substantial adverse psychological and emotional effects on patients’ social lives.3

This becomes more problematic because the lack of dermatologists and the long-awaited time for an appointment leads to patients not receiving timely acne treatments. This delay in starting the treatment results in permanent, disfiguring acne scars.3

Currently, there are numerous options available for acne scar treatment, but choosing the effective and appropriate option and its related protocols for ideal acne scar treatment highly depends on the acne scar types and subtypes.3 In addition, classifying the scar types is not straightforward. Nowadays, they are classified into three types, atrophic, hypertrophic, and keloidal, where atrophic scars are further subdivided into icepick, rolling scars, and boxcar. Resulting in a total of five acne scar classes.3

For clinical purposes, dermatologists usually classify acne scars into these five types based on a simple visual inspection, which is time-consuming, has high error rates, and causes discomfort for patients. Additionally, the description of the scars is based on the textural irregularities and the physical characteristics, such as color, shapes, patterns, width, depth, 3D architecture, etc., which makes the process of classification highly subjective to the expert’s experience and knowledge.3

Suitable treatment options need to be selected based on the type of scar. For example, atrophic acne scars can be successfully treated at home with topical over the counter retinoids. In contrast, hypertrophic and keloid scars are softened, and their height is reduced using gel silicone sheets. Hence, it is vital to have an efficient and fast automated acne classification system with highly accurate performance for timely treatment with low cost and minimal dependence on an expert’s help.3

With all this in mind, Masum Shah Junayed and colleagues, the authors of a study published in the IEEE Access journal in 2021, aimed to automate acne scar classification using artificial intelligence. To the authors' best knowledge, this is something not performed before. This is mainly because of the existence of many different human skin tones, large variations in the appearances of acne scars, such as their shape, size, and even the position of the scar. Also, acne scars depend on the patients’ age, gender, and skin type.3

Furthermore, there are no appropriate and publicly available datasets with a large number of acne scar images already labeled by a health professional that can be used by methods based on artificial intelligence. To be clearer, there are datasets for classification of acne lesions, like the study described at the beginning of this podcast, but not for acne scars. To fill this gap, the authors generated the first acne scar dataset, named ‘‘5-class Acne Scar’’, that includes 250 images labeled in five classes of acne scar types by four experienced specialists.3

Afterwards, the authors developed a method based on convolutional neural networks that performs scar classification. First, the images were preprocessed because the initial quality can greatly affect the results; thus, this step was essential to improve image quality and to have more accurate image classification. The preprocessing step involved the enhancement of the contrast to reduce the noise while preserving the edges of the acne scars. Another step in the preprocessing is increasing the number of images to avoid overfitting and improve performance, this is achieved by rotating the images randomly 30 degrees left to right and flipping the images horizontally.3

After the preprocessing, the images were fed into a novel multi-layer deep convolutional neural network model developed by the authors named “ScarNet”, which has minimal computational cost. The training and testing of the model was done using a proportion of 80 and 20 percent of the images for each set, respectively. Later, the performance of the model was evaluated, showing an accuracy of up to 92 percent and a precision of 81 percent.3

In addition, when the model is tested for specific types of acne scars, the results show that it performs best at classifying the Icepick scar types, with an accuracy of 94 percent, followed by the keloid type with 94 percent, the boxcar with 93 percent, the hypertrophic with 91 percent and the rolling type with 89 percent.3

Moreover, the authors wanted to compare the performance of their method with other types of classifiers, such as decision trees, multi-layer perceptron, support vector machine and random forest, which are machine learning classifiers. The comparison showed that the ScarNet method developed by the authors outperforms all the four other classifiers in terms of accuracy, precision, sensitivity and specificity. For instance, ScarNet achieved at least 3.4 percent more accuracy than the other models.3

As for any classification model, and despite the high-performance accuracy, there are also a few misclassification cases. When analyzing such misclassified images in detail, it is clear that the challenge in classifying the images is due to the presence of shadows, highly bright illumination, poor clarity, or low image quality. It is also possible that the model misclassified images with several different acne scar types within small skin regions simultaneously.3

This work is especially important because, when considering the extensive availability of smartphones, ScarNet could be potentially used by ordinary people as a remote screening system, especially in underdeveloped countries, where there is a lack of dermatologists.3

In conclusion, these studies demonstrate the validity of object detection with artificial intelligence on the problem of facial acne analysis, showing that they are capable of precisely locating out the acne lesions on facial images.1 This technology can also be applied to the assessment of acne scars, like the ScarNet method, which was able to accurately classify the acne scars with minimized computational cost.3

Thanks for joining us on this episode of Health Connect. Don’t miss out on our next episode. Discover more medical news and content on Viatris Connect.



Referências:

  1. 1. Hao W, Wenjian Y, Yuanqing W, Jun Z, Xiaolong L, Zhexiang K, et al. Acne Detection and Severity Evaluation with Interpretable Convolutional Neural Network Models. Technol Health Care. [Internet]. 2022. (Accessed on June 7, 2022);30:S143-S153. Available at: https://doi.org/10.1371/journal.pone.0253421
  2. 2. Razˇnatović Đurović M, Janković J, Đurović M, Spirić J, Janković S. Adolescents’ beliefs and perceptions of acne vulgaris: A cross-sectional study in Montenegrin schoolchildren. PLoS ONE. [Internet]. 2022. (Accessed on June 7, 2022);16(6): e0253421. Available at: https://doi.org/10.1371/journal.pone.0253421
  3. 3. Junayed MS, Islam MB, Jeny AA, Sadeghzadeh A, Biswas T, Shah AFMS. ScarNet: Development and Validation of a Novel Deep CNN Model for Acne Scar Classification With a New Dataset. IEEE Access. [Internet]. 2022. (Accessed on June 7, 2022);10:1245-1258. Available at: https://doi.org/10.1109/ACCESS.2021.3138021

Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.

NON-2022-12616
NON-2023-2512