Artificial Intelligence in the Diagnosis and Management of Skin Cancer

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 field. In this episode we will talk about how artificial intelligence has contributed to the diagnosis and management of patients with skin cancer.

Did you know that even though the mortality rate of skin cancer is significantly high, early detection helps to bolster the survival rate to over 95%?1

It is a challenge to estimate the incidence of skin cancer due to various reasons that include the multiple sub-types of skin cancer. As of 2020, the World Cancer Research Fund International reported a total of 300,000 cases of melanoma in skin, and more than one million cases of non-melanoma skin cancer.1

The reasons for the occurrence of skin cancer cannot be singled out, but they include and are not limited to, exposure to ultraviolet rays, family history, or a poor immune system.1

The affected spot on the skin, or lesion, can be further segregated into multiple categories depending on its origin. A comparison between different lesion types is usually accompanied by the presence or the absence of certain dermoscopic features.1

Cutaneous melanoma is the deadliest type of skin cancer; however, prevention strategies and diagnosis at early stages significantly improve survival.2

Nevertheless, the treatment of advanced melanoma remains limited, and early detection and surgical treatment still constitute the best means to improve the outcome of melanoma patients.2

The prognosis of melanoma is defined mainly through histopathological features, where Breslow thickness and the presence of ulceration are considered the most important hallmarks. Unfortunately, both these features are most relevant when the neoplastic lesion is already at an advanced stage.2

Around 20% of patients diagnosed with early-stage melanoma, stages I and II according to the American Joint Committee on Cancer, develop metastasis in the following 5 years, and, interestingly, some stage II patients have worse survival than stage III patients. This phenomenon can be explained by the heterogeneity in the changes during the evolution of this neoplastic disease, both the genetic alterations and those related to melanoma cell plasticity.2

Despite clinical staging guidelines, the heterogeneous nature of melanoma makes its diagnosis and prognosis challenging, even for experienced dermatologists. Indeed, the wide range of morphologies among skin lesions makes it difficult to distinguish melanomas from other pigmented skin lesions.2

The diagnosis of a primary melanoma is achieved by visual examination of suspicious lesions followed by biopsy. Moreover, the acquisition of images of the lesion can now be complemented by dermatoscopic examination.2

Dermatoscopy, also known as dermoscopy, is a non-invasive diagnostic method, which allows for a closer examination of the pigmented skin lesion with a dermatoscope. This procedure allows visualization of the skin structure in the epidermis that would otherwise not be possible to the naked eye.1

Dermoscopy can be categorized into three modes—polarized contact, polarized noncontact, and nonpolarized contact (or unpolarized dermoscopy). Polarized and nonpolarized dermoscopy are complementary, and utilizing both to acquire clinical images increases the diagnostic accuracy. Then, these images can be processed with the help of AI methods to assist in the diagnosis of skin cancer.1

There are three stages associated with an automated dermoscopy image analysis system, namely pre-processing, image segmentation, and feature extraction. Segmentation plays a vital role, as the succeeding steps are dependent on this stage’s output. Segmentation can be carried out in a supervised manner by considering parameters such as shapes, sizes, and colors, coupled with skin texture and type. As is the case of melanoma development that takes place horizontally or radially along the epidermis, which carries critical importance in the early diagnosis of skin cancer.1

Artificial Intelligence has laid the foundation for integrating computers into the medical field seamlessly. It provides an added dimension to diagnosis, prognosis, and therapy.1

Artificial intelligence-based methods are being developed to automate the analysis of images in order to facilitate early diagnosis of melanoma, and to help dermatologists in both hospital environments and in primary care consultations.2

Machine learning and deep learning models for skin cancer screening have been on the rise. This is primarily because these models, as well as other variants of artificial intelligence, use a concoction of algorithms, and when provided with data, accomplish tasks. Such tasks include, but are not limited to, diagnosis, prognosis, or the prediction of the status governing the ongoing treatment.1 

Diagnosis is the process of understanding the prevailing state of the patient, while prognosis refers to the process of predicting the future condition of the patient by extrapolating all the current parameters and their corresponding outputs.1 

Artificial intelligence has now progressed to the point where it can be successfully used to detect cancer earlier than the traditional methods. As early detection is key for the fruitful treatment and better outcome of skin cancer, the need for machine learning and deep learning models in the field of skin cancer is paramount.1

Among Machine Learning Techniques, Artificial Neural Networks have been used to predict non-melanoma skin cancer by inputting a certain set of tried and tested parameters fit for training, such as gender, vigorous exercise habits, hypertension, asthma, age, heart disease, among others. The Artificial Neural Network takes the entire dataset as the input. To improve the accuracy of the model, the network inputs are normalized to values between 0 and 1. The outputs are treated as typical classification outputs, which return fractional values between 0 and 1. Artificial Neural Networks can also be used to detect skin cancer by taking an image input and subjecting it through hidden layers. Therefore, Artificial Neural Networks play an important role in predicting skin cancer and the presence of a tumor, due to their flexible structure and data-driven nature, for which they are considered as a potential modeling approach.1

The development of computer-aided diagnosis tool applications in dermatology is an area of research that has achieved significant relevance in recent years due to the increasing number of appointments in primary care and at the dermatology units.2

The latest advances in artificial intelligence, and more specifically in deep learning, represent a landmark in modern medicine. Solutions based on these tools have been seen to perform at an expert level, paving the way to adopt such computer-aided diagnosis tool solutions in clinical practice, and although there are still several challenges that remain to be overcome, great progress continues to be made.2

For example, Diaz-Ramón and colleagues designed a new tool to improve the early diagnosis of melanoma by using epiluminescence dermatoscopy and deep learning image analysis. It is called the Melanoma Clinical Decision Support System, or Melanoma CDSS, which is a bioinformatics application based on artificial intelligence algorithms capable of extracting information that may be useful to improve the diagnosis and prognosis of patients with cutaneous melanoma. It makes use of clinical, demographic, and molecular data obtained by quantifying serum cytokines and other histopathological biomarkers in melanoma biopsies from patients.2

The initial evaluation of dermatological lesions, specifically melanomas, relies on the identification of ABCDE features, whereby lesion asymmetry, borders, color, diameter, and evolution are studied. Lesion features can be studied by eye or assisted by dermatoscopy, but the final diagnosis relies on biopsy and histopathological diagnosis, which takes 3 or 4 weeks to complete.2

Because of this delay in the final diagnosis, this computer-aided diagnosis tool was developed by the researchers to assist clinicians in the diagnostic procedure when examining a lesion. The Melanoma CDSS is an application capable of extracting useful information to help clinicians reach a diagnosis and define the prognosis of melanoma patients based on AI models.2

CÁPSULA

The preconceived notion surrounding the use of artificial intelligence in the cancer diagnosis domain that the introduction of technology may eventually downsize the workforce, has brought about apprehensions about adopting artificial intelligence.1

Decision support systems are computerized programs used to assist with decision making and choosing the right course of action. Artificial-Intelligence-powered decision support systems can be used in the diagnosis of skin cancer. They provide options of flexibility in designing deep learning classification models by hinting at the common procedures and looping patterns.1

Support systems can be initialized with pretrained deep neural network models combined with transfer learning to classify skin lesion localization and classification. Present day decision support systems are fused with automated deep learning methods. These methods are fine-tuned and trained with the help of transfer learning using imbalanced data. The model extracts the features using an average pooling layer, although the extracted features are not sufficient.1

Using artificial intelligence-powered decision support systems can help clinicians diagnose and potentially replace invasive diagnostic techniques.1

FIN DE LA CÁPSULA

In recent decades, new therapeutic perspectives for cancer treatment have shifted toward precision medicine, personalized to patient characteristics. The continuous discovery of new molecular markers and the use of innovative techniques make possible a more delineated view of the tumor and less harmful treatments.3

The clinical choice to perform a specific treatment involves the analysis of many factors, such as disease grade, mutational status and, of course, the patient’s condition. Such choices, although linked to standard treatment regimens, are often difficult to interpret in many cancer contexts, such as metastatic melanoma.3

Therefore, the development of AI methods capable of interpreting and analyzing a massive amount of data, could help clinicians in the management of melanoma patients and the most appropriate therapeutic choices.3

In addition, such methods could help in the prediction of possible disease recurrence and response to standard treatments, thus hypothesizing different treatment scenarios.3 

In this context, Goussalt and colleagues developed and validated four machine learning models to predict the response to immunotherapy and targeted therapy in stage IIIc or stage IV melanoma patients. The work was conducted on data from 10 centers participating in the French network for Research and Clinical Investigation on Melanoma, launched in 2012, with about 935 patients, corresponding to 1,978 systemic treatments extracted from that database.3

Several data were considered in this study, including age, sex, melanoma type, spontaneous regression, number of invaded lymph nodes, extracapsular extension, mutational status, melanoma stage, number of metastasis sites, and lines of treatments.3

Complete or partial response and stable disease were defined as class 1, while progressive disease was defined as class 2. The algorithm performances were evaluated on the test set by the percentage of treatments correctly classified in class 1 or 2.3

The authors identified and validated four predictive algorithms for drug response for both types of treatment, immunotherapy and targeted therapy. These machine learning models confirmed, for melanoma, the validity of several predictor variables of response to treatments already found in the literature.3

In recent years, machine learning has been applied extensively to improve melanoma risk stratification and prognosis prediction. Most research has focused on finding new clinical and pathological markers. Nevertheless, none of the new, promising, prognostic variables have yet been added to the American Joint Committee on Cancer system, which is currently the gold standard staging system.4 

Stage II and III patients currently have access to different therapeutic strategies, resulting in a subset of stage II patients having worse survival rates than stage III patients. A more accurate prognostic tool is needed to increase the survival of melanoma patients by preventing recurrence and providing the most appropriate follow-up regimens.4

In a study conducted by Cozzolino and colleagues, a machine learning algorithm was developed to effectively predict short-term overall mortality of patients with cutaneous malignant melanoma.4

Researchers focused on the implementation of an algorithm based on known and validated prognostic factors, with the aim of using machine learning to improve prediction capabilities and facilitate the application of this novel melanoma risk stratification tool.4

They found promising results when training a new model through machine learning. Using only routinely collected information, their best algorithm, a small Deep Neural Network, was able to attain an accuracy of 91.1% and an area under curve value of 93.3%. Comparatively, one previous study that evaluated the prognostic accuracy of the AJCC staging system 8th edition, had reported an area under curve of 74%.4

There is accumulating evidence that artificial intelligence and machine learning can assist clinicians to make better clinical decisions, or even replace human judgement.5

Both artificial intelligence and machine learning algorithms can work alongside consultant dermatologists, assisting clinicians in the diagnosis of skin cancers. If these findings could be replicated in primary care settings where there is a low prevalence of skin cancer, artificial intelligence and machine learning algorithms could have a substantial effect on diagnostic services.5

Even though these algorithms have great potential, they need to be evaluated carefully to ensure that they are accurate, effective, cost-effective, and safe enough for clinical use, and that increased access to skin lesion assessment will not add to the biopsy burden on specialist care providers or contribute to an overdiagnosis of melanoma.5

Thanks for joining us on this episode of Health Connect. Don’t miss out on our next episode. 

Discover more medical news on Viatris Connect.



Referências:

  1. 1. Melarkode N, Srinivasan K, Qaisar SM, Plawiak P. AI-Powered Diagnosis of Skin Cancer: A Contemporary Review, Open Challenges and Future Research Directions. Cancers (Basel). 2023;15(4):1183. Available at: https://www.mdpi.com/2072-6694/15/4/1183
  2. 2. Diaz-Ramón JL, Gardeazabal J, Izu RM, Garrote E, Rasero J, Apraiz A, et al. Melanoma Clinical Decision Support System: An Artificial Intelligence-Based Tool to Diagnose and Predict Disease Outcome in Early-Stage Melanoma Patients. Cancers (Basel). 2023;15(7):2174. Available at: https://www.mdpi.com/2072-6694/15/7/2174
  3. 3. Guerrisi A, Falcone I, Valenti F, Rao M, Gallo E, Ungania S, et al. Artificial Intelligence and Advanced Melanoma: Treatment Management Implications. Cells. 2022;11(24):3965. Available at: https://www.mdpi.com/2073-4409/11/24/3965
  4. 4. Cozzolino C, Buja A, Rugge M, Miatton A, Zorzi M, Vecchiato A, et al. Machine learning to predict overall short-term mortality in cutaneous melanoma. Discov Oncol. 2023;14(1):13. Available at: Machine learning to predict overall short-term mortality in cutaneous melanoma
  5. 5. Jones OT, Matin RN, van der Schaar M, Prathivadi Bhayankaram K, Ranmuthu CKI, Islam MS, et al. Artificial intelligence and machine learning algorithms for early detection of skin cancer in community and primary care settings: a systematic review. Lancet Digit Health. 2022;e466-e476. Available at: https://www.thelancet.com/journals/landig/article/PIIS2589-7500(22)00023-1/fulltext

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-2023-9467
NON-2023-2512