JMIR Publications

ABSTRACT

Background:

Studies have shown that artificial intelligence achieves similar or better performance than dermatologists in specific dermoscopic image classification tasks. It is however susceptible to the influence of confounding factors within images (e.g. skin markings), which can lead to false diagnoses of cancerous skin lesions. Image segmentation can remove lesion-adjacent confounding factors but greatly changes the image’s representation.

Objective:

Compare the performance of binary skin lesion classifiers trained on segmented or unsegmented dermoscopic images and determine if and how the segmentation process affects their skin lesion classification performance.

Methods:

Separate binary skin lesion classifiers (nevus vs. melanoma) were trained and evaluated on segmented and unsegmented dermoscopic images. For a more informative result, classifiers were trained on two distinct training datasets (HAM and ISIC). Each training run was repeated five times, resulting in an ensemble of classifiers for each training dataset and preprocessing method. Mean ensemble performance was evaluated on a multi-source test set (n=688) consisting of a holdout and an external component.

Results:

Segmented HAM classifiers showed a higher mean balanced accuracy of 75.6% ±1.1% compared to 66.7% ±3.2% (unsegmented), which was significant in four out of five runs (P<.001). ISIC classifiers showed a comparable mean balanced accuracy of 77.4% ±1.5% (segmented) and 78.3% ±1.8% (unsegmented), which was significantly different in one out of five runs (P=.004).

Conclusions:

Image segmentation does not result in an overall performance decrease and additionally causes the beneficial removal of lesion-adjacent confounding factors. Thus it is a viable option to address the negative impact confounding factors have on DL-models in dermatology. However, the segmentation step might introduce new pitfalls, which requires future investigation.