Comprehensive Summary
This paper by Vollmer et al. studies the usage of deep learning for the identification of benign melanomas. Using 200 skin lesion cases, the researchers compared the sensitivity to diagnose the skin lesion as malignant or benign. They compared only the dermatologists, the dermatologists aided by the deep learning model DL-CNN, and the results of only the deep learning model. The cases chosen for this study were intentionally difficult to diagnose and included 100 real cutaneous melanoma cases and 100 simulated melanoma cases. The results of this study showed that the DL-CNN model alone did significantly worse than the dermatologists in terms of sensitivity and specificity. However, the group of derma tologists aided by the DL-CNN model did slightly better than the group without. The results of this study are particularly notable because the information from the DL-CNN model did not lower the diagnostic ability but rather improved it by a relatively small but statistically significant amount.
Outcomes and Implications
It is important to identify melanomas early as benign lesions, because the cutaneous melanoma survival rate decreases as the melanomas grow in size. However, there is a difficulty in early prognosis sensitivity. Being too sensitive results in undue burden on the patient, and not being sensitive enough risks malignant growth. While the deep learning model did not prove to be viable as a sole means of diagnosis, it did somewhat improve the dermatologist’s diagnosis efficacy. The results of this study suggest that there is great potential for better models to be more helpful for diagnostic support in the future. Earlier melanoma detection, aided by deep learning, would result in better patient outcomes and fewer late-stage cancer patients.