Early detection of malignant melanoma is of critical prognostic importance. To improve sensitivity and specificity, a variety of other noninvasive diagnostic methods are now available in addition to sequential digital dermoscopy, including deep learning models.
In a retrospective cross-sectional study, neural networks (“convolutional neural networks” CNN) developed for image-based diagnostic classification of skin lesions were compared with the sequential monitoring strategy [1–3]. Image quartets of 59 high-risk patients, each with three nevi and one melanoma originally diagnosed on the basis of sequential changes by digital dermoscopy, were used as test material [3]. Two validated neural networks were used to evaluate image quartets at baseline and at the time of melanoma diagnosis. In addition, the baseline quartets were evaluated by 26 dermatologists. Relevant target criterion was the number of quartets with completely correct classification.
Study results and conclusion
At baseline, the networks correctly classified all lesions in 15.3% and 13.6% of the 59 quartets, respectively [3]. This corresponded to a sensitivity of 25.4% and 28.8% and a specificity of 92.7% and 75.7%, respectively. After sequential monitoring at the time of melanoma diagnosis, sensitivity improved to 44.1% and 49.2%, respectively, due to the development of additional morphologic melanoma features. Dermatologists correctly classified an average of 24 (22-27) of the 59 quartets at baseline, and you had been told that each quartet contained exactly one melanoma. In an alternative evaluation of the CNN results, a quartet was already considered correctly classified as soon as the melanoma was assigned the highest malignancy score in the quartet. Under this approach, the two CNNs correctly classified 28 (47.5%) and 22 (37.3%) of the 59 baseline quartets, respectively. The conclusion of the study authors is, on the one hand, that the neural networks studied could not replace the sequential monitoring strategy for melanoma detection and, on the other hand, dermatologists and neural network together seem to achieve better diagnostic performance in sequential melanoma detection [3].
Congress: DDG compact and practical
Literature:
- Sies K: JDDG 2021; 19(6): 842-851.
- Jutzi TB, Brinker TJ: Dtsch Arztebl 2020; 117(24): [14]; DOI: 10.3238/PersDerma.2020.06.12.03
- Winkler J, et al: Can neural networks replace sequential digital dermoscopy in high-risk patients? Dermatologie kompakt und praxisnah 18-20.02.2022, abstract volume, P015.
DERMATOLOGY PRACTICE 2022; 32(3): 38
