JMIR Publications

ABSTRACT

Background:

Automatic skin lesion recognition has shown to be effective in increasing access to reliable dermatology evaluation, however, most existing algorithms rely solely on images. Many diagnostic rules including the three-point checklist are not considered by AI algorithms, which comprise human knowledge and reflect the diagnosis process of human experts.

Objective:

We aim to develop a semi-supervised model that can not only integrate the dermoscopic features and scoring rule from the three-point checklist, but also automate the feature annotation process.

Methods:

We first trained the semi-supervised model on a small annotated dataset with disease and dermoscopic feature labels, and tried to improve the classification accuracy by integrating the three-point checklist using ranking loss function. Then utilized a large unlabeled dataset with only disease label to learn from the trained algorithm to automatically classify skin lesions and features.

Results:

After adding the three-point checklist to our model, its performance for melanoma classification improved from a mean (standard deviation) of 0.8867 (0.0191) to 0.8943 (0.0115) under five-fold cross-validation. The trained semi-supervised model can automatically detect three dermoscopic features from the three-point checklist, with best performances of 0.75 (0.7932), 0.89 (0.8752), and 0.76 (0.8444), in some cases outperforming human annotators.

Conclusions:

Our proposed semi-supervised learning framework can help with the automatic diagnosis of skin disease based on its ability to detect dermoscopic features and automate the label-annotation process. The framework also can help to combine semantic knowledge with a computer algorithm to arrive at a more accurate and more interpretable diagnostic result, which can be applied to broader-use cases.