Hyperspectral Imaging of Blood Through Fish Skin Using a Neural Network Machine Learning Approach

Summary

We propose a novel model for determining residual blood concentration in both cod fillets and whole fish using interactance hyperspectral imaging. The amount and distribution of residual blood in fish muscle is an important quality parameter. Improving the ability to assess quality rapidly, automatically, and noninvasively can strengthen the ex‐vessel fishing market, incentivizing practices that result in good quality, thereby reducing food waste and economic losses in the form of reduced quality. Our method employs a fully connected neural network operating pixelwise on hyperspectral images, trained on controlled experiments with known blood concentrations imaged with and without cod skin. The partial dependence of the neural network was estimated as a function of wavelength and corresponds with the well‐known absorption spectra of hemoglobin and water. Robust performance was demonstrated on real‐world data from a commercially available hyperspectral imaging system, and the model’s sensitivity to blood through skin was demonstrated by comparing predictions for whole fish with those for fillets with skin removed. This study underscores the value of controlled experiments in constraining flexible nonlinear models like neural networks, efficiently addressing complex issues such as light transport through fish skin. Our model extends the capability of hyperspectral imaging for automated quality assessment for fillets as well as gutted whole fish, which is highly relevant to the whitefish processing industry.