In the Journals

Avoid skin markings in dermoscopic images for analysis, training by convolutional neural networks

Melanoma probability scores increased due to standard skin markings from surgical ink and were associated with a significant reduction in the specificity of a convolutional neural network trained in dermoscopic images, according to an analysis in JAMA Dermatology.

“This study’s findings suggest that skin markings significantly interfered with the CNN’s correct diagnosis of nevi by increasing the melanoma probability scores and consequently the false-positive rate,” Julia K. Winkler, MD, of the department of dermatology at University of Heidelberg in Germany, and colleagues wrote.

The convolutional neural network (CNN) architecture was trained with more than 120,000 dermoscopic images and corresponding classification labels. For the study, three images sets were created of 130 melanocytic lesions with and without skin markings.

The CNN found an increase in the mean melanoma probability scores of the classifier in benign nevi with skin markings from 0.16 (95% CI, 0.10-0.22) to 0.54 (95% CI, 0.46-0.62). Moreover, in melanoma images, there was an increase in the mean melanoma probability scores in unmarked images from 0.94 (95% CI, 0.85-1.00) vs. images marked electronically to 1.00 (95% CI, 0.99-1.00).

The sensitivity of the CNN in the unmarked image set was 95.7% (95% CI, 79%-99.2%) and the specificity was 84.1% (95% CI, 76%-89.8%). In marked lesions, the sensitivity was 100% (95% CI, 85.7%-100%) and the specificity was 45.8% (95% CI, 36.7%-55.2%).

Skin markings at the periphery of benign nevi were associated with an increase in the melanoma probability scores that increased the false-positive rate by the CNN to around 40%, according to the researchers. Cropping the images to avoid inked skin markings significantly improved the specificity of the CNN.

The researchers said that work needs to be done to understand the limitations of a deep learning CNN to successfully apply it in broader applications. They also suggested avoiding markings in images intended for analysis and in training images for future algorithms. – by Abigail Sutton

Disclosures: Winkler reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.

Melanoma probability scores increased due to standard skin markings from surgical ink and were associated with a significant reduction in the specificity of a convolutional neural network trained in dermoscopic images, according to an analysis in JAMA Dermatology.

“This study’s findings suggest that skin markings significantly interfered with the CNN’s correct diagnosis of nevi by increasing the melanoma probability scores and consequently the false-positive rate,” Julia K. Winkler, MD, of the department of dermatology at University of Heidelberg in Germany, and colleagues wrote.

The convolutional neural network (CNN) architecture was trained with more than 120,000 dermoscopic images and corresponding classification labels. For the study, three images sets were created of 130 melanocytic lesions with and without skin markings.

The CNN found an increase in the mean melanoma probability scores of the classifier in benign nevi with skin markings from 0.16 (95% CI, 0.10-0.22) to 0.54 (95% CI, 0.46-0.62). Moreover, in melanoma images, there was an increase in the mean melanoma probability scores in unmarked images from 0.94 (95% CI, 0.85-1.00) vs. images marked electronically to 1.00 (95% CI, 0.99-1.00).

The sensitivity of the CNN in the unmarked image set was 95.7% (95% CI, 79%-99.2%) and the specificity was 84.1% (95% CI, 76%-89.8%). In marked lesions, the sensitivity was 100% (95% CI, 85.7%-100%) and the specificity was 45.8% (95% CI, 36.7%-55.2%).

Skin markings at the periphery of benign nevi were associated with an increase in the melanoma probability scores that increased the false-positive rate by the CNN to around 40%, according to the researchers. Cropping the images to avoid inked skin markings significantly improved the specificity of the CNN.

The researchers said that work needs to be done to understand the limitations of a deep learning CNN to successfully apply it in broader applications. They also suggested avoiding markings in images intended for analysis and in training images for future algorithms. – by Abigail Sutton

Disclosures: Winkler reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.