Model-less Perfusion Analysis using Deep Learning Framework

Among the various methods of analyzing pulmonary perfusion, the “with-model method” leads to large analysis errors when the model does not match the characteristics of the lesion. On the other hand, the deconvolution method of the “without-model method” is supposed to estimate the response relationship between a particular input waveform and the output waveform, i.e., the impulse response. However, the actual deconvolution method was not strictly “without-model method” because a model function was fitted in the calculation process. In this study, we propose the “Convolution Method” to estimate blood flow dynamics without using a specific model. Specifically, it “convolves the impulse response with the input waveform and updates the impulse response to minimize the error between the convolved waveform and the observed waveform. In Experiment 1, we analyzed waveforms assuming Delay and Dispersion, which have been considered difficult to analyze, and compared the conventional and proposed methods. In Experiment 2, blood flow analysis was performed on a patient with a left pulmonary artery defect. The results confirmed that the proposed method has high convergence, is independent of the input waveform, and can analyze even in the presence of vascular stenosis. It was also confirmed that it can analyze multi-input systems, and that it produces results consistent with medical findings even in patients with left pulmonary artery defects.