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

摘要

In this paper, with reference to practical applications, we investigate the accuracy of the PCA-NN method in determining the optical properties μ_a and μ_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 μ_a and μ_s' and the accuracy of μ_a and μ_s' was calculated, respectively. The results showed that for the reflectance with the same data length, the prediction errors of μ_a and μ_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.