Algorithm developed at Unicamp classifies skin lesions through image analysis
A group of researchers from Unicamp has been working on the development of software with the potential to speed up the diagnosis of melanoma skin cancer. Using artificial intelligence and deep learning, a machine learning technique using artificial neural networks, the team has already achieved an accuracy of 86% in diagnosis. Now, he is dedicated to improving results and developing applicability in the daily lives of health centers.
The hope is that in the near future, with the system installed on a cell phone and with a dermoscopic lens attached, it will be possible to quickly extract a diagnosis, explains Sandra Avila, professor at the Computing Institute who is part of the study. “The idea is that we put this inside a health center, for example, where there is no dermatologist. Often people only become aware of the lesion when it starts to grow, itch and bleed, when the cancer has probably already advanced and the chance of a cure is much lower, at 14%. In the early stages, the chance of cure is 97%”, says Sandra.
The researcher, who has been dedicated to the project since 2014, emphasizes that the idea is not to replace the diagnosis made by the doctor, but to provide support to this professional. “Artificial intelligence works as a support, as an aid, but the final decision always has to be made by the doctor”, she observes. Thus, combining technology with the knowledge of healthcare professionals can speed up the early detection of melanoma, which is the most aggressive and lethal type of skin cancer, improving the patient's life prognosis.
Machine diagnosis
The analysis carried out by the machine, explains Sandra, takes place through a public image bank. With the algorithms developed by the researchers, the computer can identify whether the lesion is benign or malignant. Currently, the bank has 23.906 photographs of different types of skin lesions. The more images, says the professor, the greater the possibility of the diagnosis being accurate, as the machine learns tothrough examples. Therefore, one of the prospects for advancing research is to be able to expand the database with images obtained in Brazilian hospitals.
The result of 86% in diagnostic accuracy, according to Alceu Bissoto, PhD student in Computer Science and Sandra's advisor, was observed through existing data, referring to injuries, in the database. “This 86% is not necessarily about data in a real situation, it is about a public set of images, of which we already know what the diagnoses are, and then we compare the performance of the solution with the real diagnosis, reaching 86%” .
Even when part of the image information is removed, the diagnosis continues to be correct in 71% of diagnoses, a rate higher than the 67% average accuracy of the assessment of 157 dermatologists. “Even when the information is removed, the result is still better than that 67%. But be careful: we don't want to say that the machine is better than the doctors. The most interesting question is to think about what the machine is learning that, even taking important information from a medical point of view, it continues to get it right”, highlights Sandra. The answer, which involves understanding what patterns the machine is creating and observing on its own, is what the researchers want to find in a few months, continuing another stage of the research.
Google awards study
For the fourth consecutive year, the study on the detection of melanoma, which began in 2014 through a partnership between Sandra and professor Eduardo do Valle, from the Faculty of Electrical Engineering at Unicamp, was one of those awarded by the Google Latin America Research Awards (Lara). The award was granted to 25 research studies in Latin America, 15 of which were Brazilian. Of these, 13 are linked to public educational institutions, three of which are from Unicamp.
