The application of Neural Networks for classifying impedance spectra established with needle electrode measurements in fat and muscle tissue

Abstract

Different electrical impedance properties in different types of tissue can be utilized in order to distinguish between several tissue types and to recognize each particular type. Electrical impedance spectra established with a needle electrode, where only the needle tip is uninsulated and active as an electrode, can be used in order to separate between different tissue types. Since the electrode surface area is small, the sensitivity volume of the electrode is also small. The sensitivity has a spherical shape and has the needle tip positioned in its center. The tissue within the sensitivity volume dominates the measured tissue impedance. The small sensitivity volume means that it is the tissue in the immediate vicinity of the needle tip which dominates the measured impedance. The small volume also gives a high probability for the tissue within the volume to be homogenous. For a given current I, an electrode with a small surface area is exposed to a higher current density J than an electrode with a larger surface area. Higher current density entails larger influence from Electrode Polarization Impedance (EPI). The impedance measured with the needle electrode contains EPI in series with the tissue impedance. The influence from EPI is also frequency dependent. Considerable influence from EPI must be expected in the measurements for frequencies up to 10 kHz – 50 kHz. Instead of just trying to minimize the influence of EPI in the impedance measurements by omitting the frequency range below 50 kHz, the tissue dependencies of the EPI are explored. Two data sets with impedance spectra are established. For each of the two data sets, a separate type of needle electrode is used in order to establish the impedance spectra by performing impedance measurements in many different locations in fat tissue and muscle tissue in a newly deceased pig. The impedance measurements for each set were performed on a separate afternoon/evening at Rikshospitalet in Oslo. 100 impedance spectra in fat tissue and 100 impedance spectra in muscle tissue in the frequency range between 10 Hz and 1 MHz were established with each of the two needle electrode types. The two data sets were analyzed with two types of Artificial Neural Networks (ANNs): One Feed-forward Neural Network (FNN) and one Recurrent Neural Network (RNN) with Long-Short Term Memory (LSTM) cells. By comparing separate analyzes of the spectra in the lower part of the frequency range between 10 Hz and 2812 Hz with analyzes of the spectra in the higher part of the frequency range between 3556 Hz and 1 MHz, it is possible to see if the EPI has a negative effect on the results in the lower part of the frequency range. The accuracy in the frequency range between 10 Hz and 2812 Hz are as good as, or almost as good as the accuracy in the frequency range between 3556 Hz and 1 MHz. The conclusion is then that if the influence from EPI is considerable in the impedance measurements for frequencies up to 10 kHz – 50 kHz, the EPI in the frequency range between 10 Hz and 2812 Hz must be tissue dependent.