ANNSESDIHM: An Artificial Neural Network based Signal Enhancement Scheme for Digital Inline Holography (DIH) Microscope

Digital inline holography (DIH) is a proven technique for cell and microparticle analysis. This technique has been used for auto characterization of microparticles. Recently, it has been used for the characterization of complete blood count (CBC) from a whole blood sample, where the comparable results between the traditional LH750 Haematology analyser and the DIH microscope system show a correlation of 0.87 for red blood cells (RBC) and 0.92 for white blood cells (WBC). However, we observed that the DIH microscope system is very sensitive to dust and other foreign microparticles from the surrounding environment. This hampers the quality of the micrograph. Moreover, the non-uniform illumination also hinders the signal quality of the images of the cell diffraction patterns. To overcome these problems, we have investigated various neural network-based signal enhancement schemes. The results show a significant improvement in the signal quality to that of the noisy version. The study shows that convolutional neural network (CNN) architecture is performing better in terms of signal improvement compared to the extreme learning machine (ELM). However, the ELM network is converging faster (within 1 epoch) compared to CNN. Also, we have studied the same with a fully connected network, which shows no improvement in the signal quality even for noises with low variance. The result demonstrates that we can use the unsupervised method, such as autoencoder (CNN) to improve the signal quality.