An artificial neural network model for accurate and efficient optical property mapping from spatial-frequency domain images

Conventional optical property mapping by spatial-frequency domain imaging (SFDI) is prone to relatively low efficiency due to the iterative nature of nonlinear curve-fitting based on light transfer model, such as diffusion approximation equation. This study aims at expediating the prediction of optical property mapping with high accuracy by developing an artificial neural network (ANN) model coupled with SFDI. A dataset was first generated using Monte Carlo simulations based on Graphic Processing Unit. The ANN training was then conducted and the model was optimized for multiple factors (learning rate, batch size and number of neurons). Experiments with simulation samples and a solid phantom were carried out to verify the performance of the ANN model for predicting optical property mapping. Results showed that the normalized mean absolute errors of absorption coefficient (mu a) and reduced scattering coefficient (mu's) were 0.18 % and 0.027 %, while the root mean square errors were as low as 0.01 and 0.14. Optical properties of the solid phantom demonstrated that the proposed ANN model retained the accuracy, and was about three orders of magnitude faster than the inverse curve-fitting model. Finally, the ANN model was tested on optical property measurement of 'Golden Delicious' apples with three different bruising levels (non-, light and severe bruising). Results indicated that the ANN model could measure apple mu a and mu's accurately and efficiently, and the measured mu's mapping was capable of detecting early bruising in apples. Contrast of bruise feature in mu's value between bruised and non-bruised tissues was significantly improved, enhancing apple bruising detection.

Automated summary

This study developed an artificial neural network model coupled with SFDI to predict optical property mapping efficiently and accurately. Experimental results showed that the model achieved high accuracy and was three orders of magnitude faster than traditional curve-fitting methods.