Supervised learning with low-quality ground truth in early diagnostics of malignant melanoma.

摘要

The label noise in AI-aided cancer diagnostics has various origins but often poses a challenge to the data analysis. Misclassified samples in the training set can lead to low accuracy of predictions. In this work, we present strategies of reducing the label noise in the context of dermatofluoroscopy (two-photon fluorescence excitation spectroscopy for early diagnosis of malignant melanoma) and support vector machines (SVMs). The experiments performed on real data set composed of 265 pigmented skin lesions confirm the hypothesis of reduced model accuracy in the presence of label noise. Relabeling and especially removing the supporting vector examples from the training set (100 skin lesions) allow for building models of very high predictive accuracy in diagnosing malignant melanoma as shown on independent data set (165 skin lesions). Furthermore, in the limit of very low data quantity, relabeling of supporting vectors and ensembling are shown to yield models that are more robust to label noise.