Wireless Capsule Endoscope Localization Using LSTM Network

In this letter, we propose a method of wireless capsule endoscope (WCE) localization using long short-term memory (LSTM) recurrent neural network. We use the ultra-wideband (UWB) and medical implant communication service (MICS) band path loss of the signals received by a set of sensor receivers as the input features for WCE localization. We apply local weighted linear regression (LWLR) for path loss smoothing minimizing the deviations caused by the human body tissue dielectric properties. Then, the LSTM network is applied on the scattered and smoothed path loss features to localize the WCE in 2443 traveling positions in the small intestine trajectory. We also analyze the impact of the number of sensor receivers, network topology, and human body dimensions on the accuracy of localization. It is observed that using LSTM network on the UWB and MICS smoothed path loss, significantly high accuracy of localization with root mean square error (RMSE) of 0.24 and 0.27 mm is obtained using only 16 sensor receivers in a computational time of 467.77 and 532.23 s, respectively.