Determination of Solution Concentrations using Photoacoustic Technology and Deep Learning: A comparison study

A medium containing different concentrations of inks can be difficult to characterize. This research studies the feasibility of an assembled multispectral photoacoustic (PA) system for this purpose using a pretrained Alexnet and Long Short-Term Memory (LSTM) network. We considered different ink concentrations, which were allowed to flow into a single translucent Poly Vinyl Chloride (PVC) hollow tube immersed in a water tank. A color-tunable Light-Emitting Diode (LED) is used as the source of illumination, while a single-point transducer is used to detect the PA signals produced following the light absorptions. The produced PA signal is collected as the illumination wavelength varied from 450 nm to 700 nm. We discovered the best performance using the Alexnet trained using data of 550 nm with 97.6 % testing accuracy. The similar but slightly lower classification accuracy of 95.8 % achieved by the LSTM method suggests the importance of choosing the right wavelength for the given purpose. We concluded that the feasibility of the system in differentiating different ink concentrations makes it a possible technique for the characterization of hemoglobin variants in blood.