Man vs Machine: Who Performs Better for Screening Melanoma?
Computers are the future, but does that apply to diagnostic medicine? When it comes to diagnosing melanoma, there is a need and interest in developing a screening procedure that can be used by laypersons and non-dermatologist physicians, as well as improving accuracy for dermatologists.
A recent study compared the use of different computer algorithms with experienced physicians to determine which method led to more accurate diagnoses. Using a database of previously diagnosed melanomas and benign nevi/lentigines, 25 teams used deep learning, a form of machine learning that uses types of algorithms that automatically identify increasingly abstract concepts present in data, to diagnose the images. For comparison, eight experienced dermatologists (readers) also examined the images and provided diagnoses. The readers looked at each dermascopic image to classify the lesion as either benign or malignant, and indicate the next step in treatment (obtaining a biopsy specimen or observation/reassurance). The study examined both the top 5 prediction models from the 25 teams, and a fusion of the top 5 algorithms into a single prediction model to compare diagnostic accuracy.
The results showed that individual computer algorithms were comparable with the reader dermatologists for diagnostic accuracy. The fusion technique significantly improved computer performance and the top-ranked fusion algorithm had higher average specificity than dermatologists. In terms of management, dermatologists had specificity similar to the top individual algorithm, but lower than the top fusion algorithm approach. However, some dermatologists had higher diagnostic performance than all individual and fusion algorithms in classification or management.
The authors conclude that in this controlled, artificial setting, computer vision systems are comparable to dermatologist diagnostic accuracy, and fusion algorithms may actually exceed dermatologist performance in classification of some cases. The authors caution that the study did not take into account clinical history of patients and therefore results can’t be applied to real-life situations. Nonetheless, computer vision shows promise in medical diagnosis.
Byline: Martha L. Sikes, MS, RPh, PA-C
Posted: February 8, 2018
Source: JAAD
Adapted from the original article.
[Image: PureSolution / Shutterstock]
