Artificial Intelligence Holds Promise in Diagnosing Melanoma

2019-11-08

The research results published in the online version of the Journal of the American Medical Association show that artificial intelligence algorithms can achieve diagnostic accuracy comparable to clinical experts in identifying melanoma in skin lesion images obtained through dermoscopy. In 2016, the US Preventive Services Task Force recommended against the widespread use of visual inspection for screening skin cancer due to unclear benefits and risks associated with this method. Recently, Cochrane conducted a series of comprehensive descriptions evaluating diagnostic methods for skin lesions, including but not limited to the use of dermoscopy, reflectance confocal microscopy, teledermatology, and computer or smartphone-based applications. After visual assessment, researchers found that these three methods were not very effective and were unable to address other potential barriers to melanoma screening, such as the cost of reflectance confocal microscopy equipment and variations in clinical experience. As a result, a prospective study was conducted in which researchers utilized an artificial intelligence algorithm called malignant tumor recognition deep integration to classify suspicious pigmented skin lesion photos and compared the results with evaluations by clinical experts. A total of 514 patients, almost all of whom were Caucasian, participated in the study. The dataset consisted of 1,550 skin lesion images, including 551 biopsy images and 999 control images. The researchers performed pathological analysis on the biopsy specimens and found that 22.7% of the cases were classified as melanoma. Traditionally, artificial intelligence algorithms were trained using published dermoscopy images. However, in this experiment, researchers further refined the training by using image subsets collected from each camera and compared the results of biopsy lesions with all lesions. The area under the receiver operating characteristic curve (AUROC) was used as one of the accuracy indicators for melanoma assessment. When evaluating lesion images, the AI algorithm achieved AUROCs ranging from 85% to 91% for biopsy lesions and from 91% to 96% for all lesions. In comparison, clinical doctors achieved an AUROC value of 77.8% for biopsy lesions. Furthermore, at a sensitivity of 100%, the algorithm had a specificity of 64.8%, while the specificity for clinical doctor evaluations was 69.8%. In the final conclusion, the authors of this study pointed out that "the results of this diagnostic test demonstrate that different types of artificial intelligence algorithms can accurately detect melanoma, similar to experts. Developing low-cost screening methods, such as AI-based services, can change the diagnostic approach for patients and improve the efficiency of the entire healthcare system."

RECENT AI NEWS

RECENT AI TOOLS