Analysis of skin hyperspectral images by machine learning methods*

Hyperspectral images (HSI) are optical images containing both spatial and spectral information about the imaged object. The current image analysis include different derivatives of the radiative transfer equation, which are either slow or inaccurate. A possible solution is implementation of machine learning methods (ML).

In this study, hyperspectral images of skin were analyzed using three different ML models: artificial neural network (ANN), convolutional neural network (CNN), and random forests (RF). Skin parameters were extracted and compared to the parameters extracted by the golden standard, i.e. inverse Adding-Doubling method (IAD).

The average MAE (mean absolute error) obtained on the simulated skin spectra, where IAD results served as the ground truth, was 0.003, 0.007 and 0.009 for RF, CNN and ANN, respectively. On the experimental data, the average MAE was approx. 0.06 for all ML methods. Time to estimate parametersfor from a single spectrum by ML methods  is 90 μs, compared to 0.4 s for AD.

We demonstrated that ML methods could be used to analyzeHSI images of biological tissues in almost real time resulting in slightly lower estimated parameter accuracy compared to IAD.

 

Funding: ARRS P1-0389, HRZZ IP-2020-02-3770, uniri-tehnic-18-15