SCIENTIFIC EDUCATIONAL CENTER science idea

The researchers analyzed scientific articles on the diagnosis of skin cancer using artificial intelligence technology and found that convolutional neural networks based on deep learning are most often used for this purpose. At the same time, the most accurate result (93% accuracy) is provided by systems based on machine learning, which makes them the most preferred diagnostic method. In addition, over the past 5 years, the accuracy of such algorithms has increased by more than 9%. The results of the study, supported by a grant from the Presidential Program of the Russian Science Foundation (RNF), are published in the journal Computers in Biology and Medicine. Skin cancer is one of the most common types of cancer, as it accounts for more than 40% of all identified cancers. Skin cancer is difficult to diagnose in the initial stages, because malignant forms of pigmentation can be confused with benign ones that all people have, for example, moles. At the same time, early diagnosis is extremely important, since in this case the survival rate of patients is about 99%. If the disease is detected at later stages, when the malignant nature of pigmentation becomes obvious (itching, ulcers or crusts appear, a heterogeneous dark color), the survival rate decreases to 27%. Basically, skin cancer is diagnosed using a dermatoscope, a device that allows you to highlight a potential neoplasm and examine it with a tenfold magnification. The accuracy of this analysis is 65-75%. Artificial intelligence systems are sometimes used to help doctors with early diagnosis: they compare a mole that they "see" in a patient with a set of tens of thousands of photos of age spots from medical databases. Convolutional neural networks are most often used to diagnose skin cancer, although they do not always demonstrate high accuracy. Part of the accuracy problem is related to the fact that not all databases have already marked images as malignant or benign, which is why there may not be enough data to train the algorithm. In addition, the photos are not standardized, which also reduces the reliability of diagnosis using artificial intelligence. Scientists from the North Caucasus Federal University (Stavropol) analyzed more than 10,000 scientific articles published from 2019 to 2023, and selected 171 articles that clearly spelled out the methodology for diagnosing cancer from photographs of age spots. Next, the authors grouped the articles according to which artificial intelligence algorithm was used. They identified five groups: machine learning algorithms, convolutional neural networks, neural network ensembles, multimodal neural networks and advanced intelligent methods. Machine learning algorithms are based on the fact that the program "trains" to recognize tumors on a set of images, where each photo is signed by a person as depicting a malignant or benign neoplasm, and then looks for patterns in new photos of tumors. Convolutional neural networks recognize images by breaking them into layers, in which you can then change the contrast, brightness, and color gamut without losing image quality. Neural network ensembles are a combination of several models that are trained separately for different operations and then combined. Multimodal neural networks simultaneously work with different types of data (text, numbers, photos), and advanced intelligent methods are based on other learning principles, for example, converting images into vectors. It turned out that in only 7% of the studies, scientists used multiclass databases, which included not only photos of pigment spots, but also biopsy results (for example, a blood test for cancer markers, for a common protein, studying the shape of cells in a skin sample taken from a patient). The authors concluded that in order to improve the accuracy of diagnosis, the database should include, in addition to these signs, information about the patient — his age, gender, skin type and anatomical location of the mole. These data are not always available, because although there are recommendations for collecting cancer biomarkers, there are no uniform standards for data sets yet. In 39% of the studies, the algorithm compared a photo with a database containing less than 1,000 images, which is 10 times less than needed for a high-quality sample. Therefore, even if the accuracy of the cancer diagnosis of the algorithm in the study itself is high, in practice, when the data of hundreds of patients will pass through the algorithm, the accuracy may be lower. The scientists also found that convolutional neural networks are most often used to diagnose skin cancer — in 39% of cases, whereas the analysis showed that the highest accuracy — 3% higher than that of convolutional neural networks — is achieved by machine learning algorithms. The authors found that over the past five years, the average accuracy of skin cancer recognition in machine learning—based models has increased by 9.2%, reaching 93%, while ensemble models have increased by only 3%. At the same time, the accuracy of multimodal neural networks fell by 9.7%, and convolutional neural networks — by 1%. The researchers also determined that artificial intelligence algorithms most often (37% of all multiclass-based studies) use the HAM10000 image database, which contains 10,000 photos of seven types of skin neoplasms in people of different nationalities. The use of this database increases the average accuracy of diagnostics using artificial intelligence: for example, over the past five years, its quality has increased by 6.9% to 92.3% on average for different algorithms. "The results we have obtained show the huge potential of automated early diagnosis of skin cancer based on artificial intelligence. However, such systems still carry ethical and legal ambiguity, as well as the problem of the lack of a large number of standardized clinical databases. Therefore, sometimes the model diagnoses biased, based on the diagnosis prevailing in the database used. As a result, it is not yet possible to generalize the diagnostic criteria using artificial intelligence. In the future, we need research that will help us understand how to implement artificial intelligence algorithms for auxiliary medical diagnostics, in particular, in order to more accurately detect skin cancer in the early stages," says Pavel Lyakhov, head of the project supported by an RNF grant, Candidate of Physical and Mathematical Sciences, Head of the Department of Mathematical Modeling of the North Caucasus State University Federal University. Photo source: Ulyana Lyakhova Information and photos provided by the press service of the Russian Science Foundation Information taken from the portal "Scientific Russia" (https://scientificrussia.ru /)