Artificial intelligence in dermatology and healthcare: An overview

Neural networks

Artificial neural networks are a subset of machine learning, based on algorithms that are designed to recognize patterns that are inspired by the human neural network [Figure 2].²⁵ An artificial neural network is structured as one input layer of neurons, one or more “hidden layers” and one output layer. The input data is processed through a large number of highly interconnected elements, which are called neurons or nodes.²⁰ Deep learning is a subset of artificial neural network with stacked neural networks composed of one input and one output layer, and more than one hidden layer [Figure 3]. Deep learning algorithms require advanced computation and very large data.^25,26

Close

Figure 2:

Basic structure of a neural network with a input layer, hidden layer and output layer

Export to PPT

Close

Figure 3:

The deep neural network differs from a basic neural network in having more than one hidden layers

Export to PPT

Deep neural networks can be divided into normal (one dimensional) or convoluted (two or three dimensions). Convolutional neural network has gained importance in medical image analysis for extracting patterns from images.²⁰ Convolutions are mathematic operations that are applied to pixel data in finding or filtering patterns.^20,21 Convolutional neural network is typically composed of three types of layers (or building blocks): convolution, pooling, and fully connected layers. The convolution and pooling layers perform feature extraction, whereas the fully connected layer maps the extracted features into final output, such as classification [Figure 4].²⁷

Close

Figure 4:

Processing of image through a convoluted neural network

Export to PPT

Each input image will pass it through a series of convolution layers with filters.^,- Each layer learns to recognize a specific feature. For example: Once the first layer has successfully recognized a feature like an edge, it is fed to the next layer which trains itself to recognize more complex patterns like a corner in an image.The pooling (or down sampling) layer is used to reduce the spatial dimensions to gain computational performance As the model is repeatedly trained, individual convolutions begin to identify a specific portion of the image. Hundreds of these classifiers can be linked together to identify more complex structures within each image.^20,21,27

Support vector machine

Support vector machine is a classification and regression algorithm for data classification in two classes that uses machine learning to maximize predictive accuracy while avoiding overfitting of data. It gives accuracy comparable to sophisticated neural networks when it is used in image analysis.^28,29 The support vector machine classifier is constructed by projecting training data into a higher dimensional space known as a hyperplane, which maximizes the separation between the classes [Figure 5].³⁰

Close

Figure 5:

Support vector machine algorithm

Export to PPT

Natural language processing

Natural language processing is a machine learning program concerned with the use of software programming to understand and manipulate natural language text or speech for practical purposes.² It is used to extract information from unstructured data such as electronic health records electronic health records, which includes physical examination, clinical laboratory reports, operative notes and discharge summaries or medical journals. Data entry tasks of electronic health records takes time away from patient care and is contributing largely to physician burnout. The natural language processing procedures target at turning texts to machine-readable structured data, which can then be analysed by other machine learning techniques.^26,31,32

Artificial intelligence in dermatology

Computer-based analysis of image assists in overcoming the subjective inter-observer and intra-observer variation thereby allowing an objective evaluation of parameters.⁷⁸ There is an upward trend in the use of artificial intelligence as a diagnostic aid in dermatology. The application of computational methods is being used in dermatology for faster data processing to give better and more reliable diagnoses.^79,80

This development began in the 1980s, when a greater public awareness was noted following an increased incidence of malignant melanoma, resulting in newer diagnostic modalities for early detection of lesions. Around the same time, Wilhelm Stolz and others in the Munich imaging group developed a hand-held dermatoscope for cutaneous surface microscopy.⁸¹ Later automated diagnosis of melanomas was tested using computer-aided diagnosis to reproduce the decision of the dermatologist when observing images of pigmented skin lesions either by photography, dermoscopy, and spectrophotometry.⁸²

Several image processing software in the C programming language were written with the involvement of the dermatology digital imaging group. The focus was on automatic detection of lesional borders, malignant feature detection, detection of lesion changes, and algorithmic separation of benign from malignant lesions.^81,82 In 1987, Cascinelli et al. studied the automated classification of pigmented skin lesions based on images.^79,83 Over the last three decades, there is extensive literature on the use of machine learning models for analysing and classifying the data from skin lesions.⁸⁴ It is also believed that an exponential expansion of smartphone technology worldwide will provide low-cost universal access to this vital diagnostic care.²⁶

The general steps involved in computer-aided diagnosis include:

1. Image acquisition:

   The two predominant dermatological image types are clinical and dermoscopic images with the increased use of the latter²³

2. Image pre-processing: The term pre-processing in image diagnostic procedures usually encompasses artefact removal and lesion image enhancement.²⁵ Good performance at this stage ensures the correct behaviour of the algorithms in the following stages of analysis.⁸²

   Artefact correction involves algorithms for removal of hair artefacts, dermoscopic gels, thin blood vessels, shadows, ruler markings, specular reflections and air bubbles, which can confuse diagnosis and impede the achievement of better accuracy in the automated diagnosis process.^23,82,85 Out of these, hair shafts and ruler markings are the most commonly reported artefacts.^23,85,86 Various machine learning models are used for this purpose. Filtering is a popular method to smooth a lesion image before detecting artefacts.²³

   Under image enhancement, the most important operation is colour calibration. This operation consists of recovering real colours of a photographed lesion, illumination correction, and contrast and edge enhancement.⁸² Commonly used colour spaces include red-green-blue and International Commission on Illumination among others.^23,82

3. Image segmentation: Precise border detection (segmentation) is one of the most important and crucial areas in the automated analysis of pigmented skin lesions. It is also believed to be the most difficult task due to low-contrasts surrounding the skin, fuzzy borders, the existence of artefacts and irregular structures characterizing lesional images^23,82

   Various machine learning approaches have been described for the segmentation of lesions. They are based on the colour information of the lesion, luminance, and texture. They can be broadly classified as thresholding, edge-based, fuzzy c-means, gradient vector flow snakes or region-based methods. Thresholding method is the most commonly used one and achieves good results when there is good contrast between the lesion and the skin.^23,82,87

4. Feature extraction: The extraction of specific features in a given lesion image is an essential step towards effective automated lesion image classification. The primary objective of feature extraction is to quantify the image by a set of finite numerical features.²³ The various feature descriptors studied in the algorithm for melanoma analysis are based on dermoscopic (ABCD or pattern analysis) or clinical findings (ABCDE)^88,89

   A combination of photometric features like colour, islands of colour, colour homogeneity, colour histogram etc. and textural aspects yield good results in identifying pigmented skin lesions.^23,90 The challenges in this process lie in the vast variety of images, body location, subject parameters (age), imaging parameters (lightening or camera), and the direction from which the lesion image is viewed.²³

5. Classification of lesions: Lesion classification is the last step in the process for the computerized analysis of images. Image classification involves using selected features of an image to classify pixels of the image into one of the several classes depending on specific knowledge domain.

The two main classification types as reported in the literature in relation to medical imaging are supervised classification and unsupervised classification (already discussed). The literature has reported the application of several classification methods for lesion images. Frequently used among these methods are the artificial neural network, support vector machine, decision trees, k-nearest neighbour, regression analysis classifiers, bayesian classifiers and fuzzy logic.^20,21,23,82

Pigmentary skin lesions and malignancy

Many research articles have been published classifying nonmelanoma skin cancers vs. benign and pre-malignant lesions with varying efficacy between different artificial intelligence systems.^96,97 Many systems are commercially available for computer-aided diagnosis of pigmented skin lesions which are mainly based on dermoscopy. e.g. DANAOS expert systems, DBDermo-Mips, MoleAnalyser expert systems.⁸² Multiple smartphone based apps like SkinVision, DermaAId, Skin 10, MoleScope are available for screening skin cancers, tracking changes in moles on the skin and identification of basic skin lesions.⁹⁸ Studies done on the SkinVision app scored an 80% sensitivity and 78% specificity in detecting premalignant conditions.⁹⁹

Psoriasis

Artificial intelligence has been working on many aspects of psoriasis. Various computer aided diagnostic systems have been designed for image classification and psoriasis risk stratification.^91,100 Also machine learning prediction models have been designed to determine the treatment response of psoriasis to biologics and to differentiate psoriasis from psoriatic arthritis using genetic markers.^101,102 Correa da Rosa et al. showed that the gene-expression profiles of psoriasis skin lesions, taken in the first 4 weeks on patients who are on treatment with a biological agent, can be used to accurately predict (>80% area under the ROC curve) the clinical endpoint at 12 weeks using machine learning techniques thereby reducing the assessment gap by 2 months.¹⁰³

Emam et al. studied whether machine learning could aid in predicting long-term responses to biologics in psoriasis through analysis of data of 681 psoriasis patients from the Danish registry cohort using various modelling techniques. Patients with early diagnosis and early initiation of treatment, without psoriatic arthritis, had 90% chance of continuing treatment as per the study.¹⁰⁴ Foulkes et al. noted that signals of response to therapy in patients with severe psoriasis treated with the etanercept may be systemically detectable in lesional skin, non lesional skin, and blood at baseline, before the commencement of therapy.¹⁰⁵ Automated diagnosis of other erythemato-squamous diseases such as seborrheic dermatitis, atopic dermatitis, lichen planus, pityriasis rosea and pityriasis rubra pilaris has been studied using various clinical and histopathological features.^106,107

Acne

New technologies in imaging and software solutions have been developed in acne and rosacea evaluation. A study by Min et al. showed that compared with manual counting performed by an expert dermatologist, the sensitivity and positive predictive value of the lesion-counting program was greater than 70% for papules, nodules, pustules, and whitehead comedones.^108,109

Autoimmune disorders

Machine learning has been used in various aspects of patient identification, risk prediction, diagnosis, disease subtype classification, disease progression and outcome and monitoring and management of autoimmune disorders like systemic lupus erythematosus, systemic sclerosis, vitiligo, psoriatic arthritis, rheumatoid arthritis, and systemic vasculitis.¹¹⁰

Also, machine learning techniques have been used for automated or semi-automated classification of myositis using ultrasound images. The study differentiated between normal muscle, dermatomyositis, polymyositis, and inclusion body myositis with accuracies above 70%.^97,111

Allergic contact dermatitis

Contact dermatitis is a common immunological reaction to sensitizers. The predictive screenings for the sensitizing chemicals have been usually performed using animal models such as the murine local lymph node assay. Animal models for sensitization assessment are being replaced by non-animal testing methods like Genomic Allergen Rapid Detection assay, which is based on measurements of transcriptional levels of genomic biomarkers. Artificial intelligence algorithms were used to study the changes induced in these genomic biomarkers as compared to profiles of reference chemicals on cellular stimulation with unknown sensitizers.^112,113

Ulcer evaluation and prediction

The common risk factors for chronic wounds in India are diabetes, tuberculosis, leprosy, vascular disorders, pressure ulcer, and trauma. Automated analysis has been used in various aspects of ulcer assessment like analysis of wound perimeter, surface, depth, area determination, wound delineation and composition thereby providing an objective and quantitative assessment of healing rate during treatment.^114-117 Artificial intelligence applications are also used in predicting the development of pressure injuries among surgical critical care patients.¹¹⁸

Robotics

Robots represent the physical aspect of artificial intelligence. In healthcare, Robot-assisted biopsies were one of the first uses of surgical robots and have been used for prostate, lung, breast and stereotactic brain biopsies.^119-122 Automated device for performing skin biopsies has been developed by scientists. Wherein the biopsy can be performed with fewer instruments within a shorter duration of time without the necessity of local anaesthesia.^123,124 Robot-assisted automatic laser hair removal system has been developed to automatically detect any arbitrary shape of the desired treatment area and to provide uniform laser irradiation to the designated skin area.¹²⁵ Studies on the laser hair removal system system provides consistent irradiation not only in terms of uniform distance and also the number of irradiation shots.¹²⁶ A robotic system for hair restoration (the ARTAS system) has been developed for the follicular unit extraction type of hair restoration surgery. It consists of a robotic system with a stereoscopic camera with high resolution to identify hair follicles for harvesting and implantation.^127,128

Data

Artificial intelligence applications are only as robust as the data on which they are trained. Deep learning neural networks require large amounts of data. This can be a drawback when artificial intelligence is attempted on a disease with low prevalence or when data is generalised across different populations. The heterogeneity of medical data across institutions and the complexity of the neural networks can lead to over fitting models.^4,37,40 In addition, the quality of the information extracted, is still dependent on the accuracy of the input data being entered and the infrastructure available for data sharing. Also, algorithms can underperform in conditions where there is no precedence like new side effects of drugs or treatment resistance.^1,40,

Acceptance

Detailed history and thorough clinical examination complemented by relevant investigations is the basis of diagnosis in most dermatoses. Histopathology still remains the gold standard diagnostic investigation as compared to other noninvasive investigations including artificial intelligence. Also patient care is not restricted to diagnosis but requires a holistic approach with human touch which cannot be replaced by algorithm. All these factors may challenge the acceptance of artificial intelligence in dermatology.^136,137

Reasoning (the black box problem)

In medical practice, it has traditionally been essential in clinical decision-making to know the rationale for each decision. In contrast, DL utilizes unstructured input data, and the bulk of output generation occurs within the hidden layers. It thus becomes difficult to determine which specific feature or calculation of the input data contributed to the resultant output.^40,137

Legal liability and ethics

The efficacy of the algorithms depends on large data and certain data may infringe on patients’ privacy. Therefore, it is important to create standard ethical guidelines wherein artificial intelligence can be applied and where they are mandatory. Also, in the case of an adverse event, accountability is still a grey area that has to be addressed.^4,137,138

Bias

Algorithms may also inherit the bias of the programmer or the self-learning algorithms may learn to be biased due to lack of diversity in the training material. As most machine learning programs are based on patient data which are collected from fair skinned populations, they may underperform on images of lesions in the skin of colour.^4,13,139

Job threat

There is a huge debate among experts whether or not artificial intelligence will result in large scale job losses. Some believe that Artificial Intelligence is set to take over 47% of the U.S. employment market within 20 years across all sections.^140,141 Since medical imaging is an important field associated with artificial intelligence, some experts fear a job threat for these sub-specialities, while others believe that there is no immediate threat of job loss, but working with artificial intelligence systems will enhance diagnostic efficacy and benefit patient care.^131,142,143

Limitations of the article

The article focuses on a basic understanding of artificial intelligence in healthcare and its applications in dermatology and is limited by the lack of in-depth computer knowledge of the authors. Furthermore, the applications and implications of artificial intelligence in healthcare are vast and ever-expanding. Therefore, these cannot be entirely covered in such a brief review.