Evaluation of Polysaccharide Content in Shiitake Culinary-Medicinal Mushroom, Lentinula edodes (Agaricomycetes), via Near-Infrared Spectroscopy Integrated with Deep Learning

Polysaccharide is one of the bioactive ingredients extracted from the fruiting body of Lentinula edodes (=L. edodes), which has many medicinal functions. While the content of polysaccharide can be measured by near-infrared (NIR) spectroscopy, the NIR analytical models established previously only covered L. edodes from very limited sources, and thus could not achieve high accuracy for large samples from more varied sources. Strictly, there is a nonlinear relationship between NIR spectral data and chemical label values, and traditional modeling methods for NIR data analysis have problems such as insufficient feature learning ability and difficulty in training. The deep learning model has excellent nonlinear modeling ability and generalization capacity, which is very suitable for analyzing larger samples. In this study, we constructed a novel framework with deep learning techniques on the NIR analysis of the content of polysaccharide in L. edodes. The siPLS model was established based on the combination of the bands 4797-3995 cm-1 and 6401-5600 cm-1, while the one-dimensional convolutional neural network (1D-CNN) model was established with improved feature in the treatment of the spectral data. The comparative experimental results showed that the 1D-CNN model (R2pre = 95.50%; RMSEP =0.1875) outperformed the siPLS model (R2pre = 87.89%, RMSEP = 0.6221). As such, this work has demonstrated that NIR spectroscopy with the integration of deep learning can provide more accurate quantification of polysaccharide in L. edodes. Such method can be very useful for nutritional grading and quality control of diverse L. edodes in the market.