Model-Free Multi-Input Perfusion Using Deep Learning Algorithm

Abstract

Perfusion-based lung blood flow analysis can be divided into two types: with and without a model. While the model-based approach provides physiologically accurate results, the conditions are strict and difficult to handle. On the other hand, the model-free approach is simple, but it is limited to single-input analysis where the impulse response representing the system's properties is solved from input-output relationships. In this study, a model-free method that combines simplicity and accuracy was proposed to enable analysis of multiple-input systems and aimed to standardize analysis. In the proposed method, the impulse response was formulated in a forward problem using a deep learning algorithm and directly estimated, enabling multiple-input analysis. The results of comparative experiments showed that while the proposed method is susceptible to noise, it is easy to implement and has high convergence in the range of actual SNR. However, for multiple-input analysis, since there is no model, the blood flow components interfere with each other, causing a decrease in accuracy.