A comparison between hyperspectral imaging and RGB-based tissue perfusion assessment of human skin using...

The aim of this study was to investigate whether perfusion indices calculated from hyperspectral imaging (HSI) could be determined from RGB (red-green-blue) images using convolutional neural networks (CNNs). Furthermore, CNNs were trained on HSI-images to evaluate the capability of the CNNs in predicting the perfusion indices.

HSI images were taken from the skin of 34 volunteers in the spectral range 500-1000 nm using a CE-certified HSI device. The perfusion indices Near Infrared Perfusion Index (NIR), Superficial Tissue Oxygenation (StO2), Tissue Haemoglobin Index (THI) and Tissue Water Index (TWI), were calculated from the HSI images in the numerical range from 0 to 100 at the pixel level and divided into 8 and 32 perfusion index-based classes (PIBC). CNNs were trained on the RGB-images (3 channels) and HSI-images (100 channels) to predict PIBC in both gradations (8 and 32 classes). Prediction accuracies were assessed on separate test data at the pixel level. A mean prediction error (MPE) was introduced to assess the prediction error in incorrectly predicted pixels.

The HSI-based models generally showed higher prediction accuracies than the RGB-based models. Furthermore, the 8-class PIBC models were nominally more accurate than the corresponding 32-class-PIBC models. The MPE obtained indicates that misidentified pixels were predominantly predicted as an adjacent class to the ground truth.

CNNs are generally suitable for determining the perfusion indices. However, none of the four parameters could be reliably determined from RGB images. Increasing the number of ground truth and improving the models could improve prediction accuracy. The introduced MPE could be useful as a performance measure for models trained to assess tissue status.