A Deep Learning Approach for Separating Gamma-Ultrasound Signals using Autoencoder

Recently, the development of Gamma-Ultrasound (γ-US) probes has gained significant attention for their potential to simultaneously acquire anatomical and molecular information. However, separating gamma and ultrasound signals acquired in one readout channel has been a challenging task, mainly due to the limitation in signal bandwidth. To address this issue, we propose a deep learning-based Autoencoder model that can distinguish between gamma and ultrasound signals regardless of their frequency bands. We evaluated our proposed method and compared it with the Finite Impulse Response (FIR) filter method by varying the frequency of the gamma signal and US signal. The FIR filter method caused a significant reduction (up to 48.76%) in the peak voltage of the gamma signal as the signal length decreased, while our Autoencoder model showed similar performance to the original signal with only a negligible decrease (0.17%). Our findings demonstrate the potential of the proposed Autoencoder method to overcome the limitations of FIR filters and enable the application of Gamma-Ultrasound probes in various fields.