The region-growing algorithm had the best segmentation performance in an assessment of the effectiveness of artificial intelligence methods for melanoma classification, according to findings published in Journal of the American Academy of Dermatology.
“The region-growing segmentation algorithm incorporated with the logistic regression classifier can achieve a better performance compared with the other algorithms in this review,” Xiaoyu Cui, PhD, of the department of biomedical informatics at Northeastern University in Shenyang, China, and colleagues wrote.
Using a data set of 2,200 dermoscopic images, researchers sought to identify the best AI based on four steps: image preprocessing, image segmentation, feature extraction and classification.
Images from the International Society for Digital Imaging of the Skin were used, of which 564 were melanomas and 1,636 nonmelanomas and a subdata set included 606 images made up of 295 melanomas and 311 nonmelanomas.
Color images were converted to grayscale and filtered to reduce random noise, according to the researchers. Images were cropped to remove any edges and enhanced for contrast.
After preprocessing, four entirely automatic image segmentation algorithms were implemented: active contour model, cluster, region growing and Otsu.
Researchers assessed the accuracy, specificity and sensitivity of the algorithms.
Within the small data set of 606 images, region growing achieved the best segmentation results, with a final average intersection over union of 70.06% and false-positive rate of 17.67%.
The sensitivity of the logistic regression algorithm was 76% and the specificity was 87% and was deemed more suitable for the classification of melanoma dermoscopic images over the support vector machine, which had a sensitivity of only 71%.
AlexNet, Visual Geometry Group network 16 (VGG16), VGG19 and Inception V3 (Google) were used for the classification experiments. The Inception V3 model results were deemed the best with an accuracy rate of 94%, sensitivity of 95.3% and specificity of 92.1%.
“As the depth of the network increases, the accuracy of the experimental classification also increases and the time consumed also increases,” the researchers wrote.
In-line with the clinical diagnostics of melanoma, color and texture features play an important role in distinguishing melanoma with other kinds of pigment, according to researchers.
As for study limitations, the melanoma samples used were not divided by severity and all of the nonmelanomas were benign nevi.
“Because of the limited amount of medical data and various quality issues, incorporating expert knowledge into the deep learning process is an important research topic,” Cui and colleagues wrote. – by Abigail Sutton
Disclosures: The authors report no relevant financial disclosures.