Dermatologists and two Convolutional Neural Networks (CNNs) showed only fair to moderate accuracy in predicting melanoma thickness based on dermatoscopic images, according to the results of a new study.
Now in-print in the Journal of the European Academy of Dermatology and Venereology, the paper from the University of Gothenburg is based on 1,456 dermoscopy images of melanoma assessed by 438 dermatologists and two CNNs, one of which was a machine-learning algorithm trained in classifying melanoma depth. The images included 788 cases of melanoma in situ (MIS), and 474 melanomas ≤1.0 mm and 194 >1.0 mm.
This open, web-based, international, diagnostic study used an online platform from May 10, 2021 to July 19, 2021. The researchers wanted to evaluate how accurately a large number of international dermatologists were able to discriminate between MIS and invasive melanomas, as well as estimate the Breslow thickness of invasive melanomas pre-operatively, based on the dermatoscopy images.
Among the dermatologists, overall accuracy was 63% for correct classification of MIS, and 71% for invasive melanomas. According to the study authors, dermatologists outperformed a de novo CNN but not a pretrained CNN in differentiating MIS from invasive melanoma.
A total of 22,314 images were analyzed. Results showed the overall accuracy (95% confidence interval) for predicting melanoma lesion thickness was 56.4% (55.7-57.0%), 63.4% (62.5-64.2%) for diagnosing MIS, and 71.0% (70.3-72.1%) for diagnosing invasive melanoma. The accurate prediction of Breslow thickness was higher in lesions ≤1.0 mm (including MIS) (85.9% [85.4-86.4%] than in melanomas >1.0 mm (70.8% [69.2-72.5%]).
“As well as providing valuable prognostic information, the thickness may affect the choice of surgical margins for the first operation and how promptly it needs to be performed,” the study’s first author, Sam Polesie, said in a press release. He is associate professor (docent) of dermatology and venereology at Sahlgrenska Academy, University of Gothenburg, and a dermatologist at Sahlgrenska University Hospital.
“Our study highlights the difficulties of correctly assessing melanoma thickness on the basis of dermoscopic images,” he said. “In future studies . . . we want to test whether clinical decision-making in this situation can be improved by means of machine-learning algorithms.”