AI and Dermatologists Similarly Judge Melanoma Thickness
Readers most accurately discriminated between thin (≤1.0 mm) and thick melanomas (>1.0 mm)
By Physician’s Briefing Staff | July 26, 2022
Both readers and artificial intelligence predict melanoma thickness with fair to moderate accuracy using dermoscopy images, according to a study published online July 16 in the Journal of the European Academy of Dermatology and Venereology .
Sam Polesie, M.D., from the Sahlgrenska Academy at the University of Gothenburg in Sweden, and colleagues evaluated how accurately an international group of readers could discriminate between melanoma in situ (MIS) and invasive melanomas and estimate the Breslow thickness of invasive melanomas based on dermoscopy images. The analysis included 22,314 readings by 438 international readers.
The researchers found that the overall accuracy for distinguishing melanoma thickness was 56.4 percent; the overall accuracy rates for correctly classifying MIS and invasive melanoma were 63.4 and 71.0 percent, respectively. For melanomas ≤1.0 mm (including MIS), readers accurately predicted the thickness in 85.9 percent of images compared with 70.8 percent of melanomas >1.0 mm. For differentiating MIS from invasive melanoma, the reader collective outperformed a de novo convolutional neural network but not a pretrained convolutional neural network.
"Our study highlights the difficulties of correctly assessing melanoma thickness on the basis of dermoscopic images," Polesie said in a statement. "In future studies, we aim to explore the usefulness of predefined dermoscopic structures for distinguishing. We also want to test whether clinical decision-making in this situation can be improved by means of machine learning algorithms."
One author disclosed financial ties to the pharmaceutical industry.