Purpose: Current skin cancer detection relies on dermatologists’ visual assessments of moles directly or dermoscopically. Our goal is to show that our similarity assessment algorithm on dermoscopic images can perform as well as a dermatologist’s assessment.
Approach: Given one target mole and two other moles from the same patient, our model determines which mole is more similar to the target mole. Similarity was quantified as the Euclidean distance in a feature space designed to capture mole properties such as size, shape, and color. We tested our model on 18 patients, each of whom had at least five moles, and compared the model assessments of mole similarity with that of three dermatologists. Fleiss’ Kappa agreement coefficients and iteration tests were used to evaluate the agreement in similarity assessment among dermatologists and our model.
Results: With the selected features of size, entropy (color variation), and cluster prominence (asymmetry), our algorithm’s similarity assessments agreed moderately with the similarity assessments of dermatologists. The mean Kappa of 1000 iteration tests was 0.49 (confidence interval ( CI ) = [ 0.23 , 0.74 ] ) when comparing three dermatologists and our model, which is comparable to the agreement in similarity assessment among the dermatologists themselves (the mean Kappa of 1000 iteration tests for three dermatologists was 0.48, CI = [ 0.19 , 0.77 ] .) By contrast, the mean Kappa was 0.22 (CI = [ − 0.00 , 0.43 ] ) when comparing the similarity assessments of three dermatologists and random guesses.
Conclusions: Our study showed that our image feature-engineering-based algorithm can effectively assess the similarity of moles as dermatologists do. Such a similarity assessment could serve as the foundation for computer-assisted intra-patient evaluation of moles.