Optical parameters inversion of tissue using spatially resolved diffuse reflection imaging combined with LSTM-attention network

Authors: SUN, DANNI - Jiangnan University; WANG, XIN - Jiangnan University; HUANG, MIN - Jiangnan University; ZHU, QIBING - Jiangnan University; Qin, Jianwei - Tony Qin

Submitted to: Optics Express
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 2/20/2023
Publication Date: 3/6/2023
Citation: Sun, D., Wang, X., Huang, M., Zhu, Q., Qin, J. 2023. Optical parameters inversion of tissue using spatially resolved diffuse reflection imaging combined with LSTM-attention network. Optics Express. 31(6):10260-10272. https://doi.org/10.1364/OE.485235.
DOI: https://doi.org/10.1364/OE.485235

Interpretive Summary: Knowledge of optical properties of food is critical to develop sensing techniques for internal quality and safety evaluation. Spatially resolved diffuse reflectance imaging combined with inverse algorithms is an effective method for nondestructive determination of the optical properties (i.e., absorption and scattering coefficients). Traditional data-driven inverse models treat spatially resolved light distribution patterns as a disordered vector, which may lead to the risk of losing effective information. This study proposed a novel inverse algorithm based on long and short time memory networks and attention mechanisms. The LSTM-attention network divided the spatial distribution curve of photons into multiple sub-intervals using a sliding window technique, and took the sub-intervals as the input to the LSTM model. Then an attention mechanism was used to fuse the output of each module to achieve a higher accuracy for estimating the absorption and scattering coefficients. The model was built using Monte Carlo simulation data and validated by the hyperspectral scattering images of liquid samples. This study provides a new method to improve the accuracy of measuring the optical properties of turbid biological materials, which would benefit researchers who are interested in evaluating internal quality and safety attributes of food and agricultural products.

Technical Abstract: Steady-state spatial resolution diffuse reflectance imaging (SR) combined with inversion model is an effective method to realize nondestructive measurement of optical parameters, including absorption coefficient (µa) and reduced scattering coefficient (µs'), in agricultural products. Existing optical parameter inversion models based on SR can be divided into mechanism model and data-driven model. Compared with the mechanism model with higher time complexity, the data-driven model based on learning the spatial distribution pattern of photons on the sample surface of a large number of known tissue optical parameters can realize fast estimation of optical parameters. However, traditional data-driven models treat the model input (light intensity distribution pattern) as a disordered vector, which may lead to the risk of losing effective information. In this study, a novel optical parameter inversion method based on long and short time memory network and attention mechanism (LSTM-attention network) is proposed. The proposed LSTM-attention network divides the spatial distribution curve of photons into multiple consecutive and partially overlapping sub-intervals using the sliding window technique, and takes the divided sub-intervals as the input of the LSTM module, and then introduces an attention mechanism to fuse the output of each module to obtain a higher accuracy estimate of the optical parameters. In order to meet the data-driven model's need for a large number of labeled samples, the proposed LSTM-attention model is based entirely on Monte Carlo (MC) simulation data. Experimental results for test set generated by MC simulation showed that the mean relative errors (MRE) of µa and µs' were 5.59% and 1.18%, respectively, which are significantly better than the three comparative models. The study further validated the performance of the LSTM-attention model using hyperspectral scattering images of 36 liquid simulation samples prepared by three different absorbers (black ink, blue dye and green dye), covering the wavelength range of 530 nm-900 nm. The test results of the simulation samples showed that the LSTM-attention model developed by MC data achieved the best performance, with 14.89% of µa and 9.76% of µs' in MRE indictor. Therefore, the LSTM-attention method provides an effective means to improve the estimation accuracy of optical parameters for agricultural products.