Adipocyte size evaluation based on photoacoustic spectral analysis combined with deep learning method

Abstract

Adipocyte size, i.e., the cell area of adipose tissue, is correlated directly with metabolic disease risk in obese humans. This study proposes an approach of processing the photoacoustic (PA) signal power spectrum using a deep learning method to evaluate adipocyte size in human adipose tissue. This approach has the potential to provide noninvasive assessment of adipose tissue dysfunction, replacing traditional invasive methods of evaluating adipose tissue via biopsy and histopathology. A deep neural network with fully connected layers was used to fit the relationship between PA spectrum and average adipocyte size. Experiments on human adipose tissue specimens were performed, and the optimal parameters of the deep learning method were applied to establish the relationship between the PA spectrum and average adipocyte size. By studying different spectral bands in the entire spectral range using the deep network, a spectral band mostly sensitive to the adipocyte size was identified. A method of combining all frequency components of PA spectrum was tested to achieve a more accurate evaluation.