Accuracy of PCA-NN in non-invasively determining tissue optical properties from spatially resolved diffuse reflectance

In this paper, with reference to practical applications, we investigate the accuracy of the PCA-NN method in determining the optical properties mu(a) and mu(s)' from the spatially resolved relative reflectance data produced by Monte Carlo simulations. To test prediction performance of PCA-NN from the reflectance data with different lengths and different measurement noises, we constructed six PCA-NNs respectively corresponding to data length = 5, 10, 15, 20, 25 and 30 mm, which were trained by higher precision reflectance produced with photons = 10(7). Then lower precision reflectance generated with photons = 10(4), 2 x 10(4), 5 x 10(4), 7 x 10(4), 10(5), 2 x 10(5), 5 x 10(5), 7 x 10(5) and 10(6) were inputted to PCA-NNs to extract mu(a) and mu(s)' and the accuracy of mu(a) and mu(s)' was calculated, respectively. The results showed that for the reflectance with the same data length, the prediction errors of mu(a) and mu(s)' increase as the data noise increases; but for the reflectance with the same data precision, the errors decrease as the data length becomes longer. In conclusion, the preliminary results in this paper provide a guideline for choosing appropriate measurement conditions or estimating the prediction errors in reality.