Machine learning-based classification of spatially resolved diffuse reflectance and autofluorescence spectra acquired on human skin for actinic keratoses and skin carcinoma diagnostics aid.

Significance: The incidence of keratinocyte carcinomas (KCs) is increasing every year, making the task of developing new methods for KC early diagnosis of utmost medical and economical importance.

Aim: We aim to evaluate the KC diagnostic aid performance of an optical spectroscopy device associated with a machine-learning classification method.

Approach: We present the classification performance of autofluorescence and diffuse reflectance optical spectra obtained in vivo from 131 patients on four histological classes: basal cell carcinoma (BCC), squamous cell carcinoma (SCC), actinic keratosis (AK), and healthy (H) skin. Classification accuracies obtained by support vector machine, discriminant analysis, and multilayer perceptron in binary- and multi-class modes were compared to define the best classification pipeline.

Results: The accuracy of binary classification tests was $>80\%$ to discriminate BCC or SCC from H. For AK versus other classes, the classification achieved a 65% to 75% accuracy. In multiclass (three or four classes) classification modes, accuracy reached 57%. Fusion of decisions increased classification accuracies (up to 10 percentage point-increase), proving the interest of multimodal spectroscopy compared with a single modality.

Conclusions: Such levels of classification accuracy are promising as they are comparable to those obtained by general practitioners in KC screening.