Feasibility of Deep Convolution Neural Network-Based Automatic Time Activity Curve Fitting Method for Non-Invasive Cerebral Blood Flow Quantification

Abstract

In this study, we aimed to develop a deep convolutional neural network (DCNN)-based automatic timeactivity curve (TAC) fitting method for input function determination. This will be achieved through a comparison between the DCNN method, manual method, and mathematical fitting methods using the expectation maximization algorithm (EM-method) to uncover the potential of the DCNN approach.A U-Net architecture based on convolutional neural networks was used to determine the TAC fittings. The area under the curve (AUC) values of the TAC by the EM and DCNN methods were compared to those obtained using the manual method.The AUC values for the EM-manual method exhibited similarity within an error range of approximately ±20%. Conversely, the error range for DCNN-manual method was approximately ±10%, signifying a reduction in the error range to approximately 1/2.Our findings indicate that the DCNN method provides accuracy equivalent to those of manual methods and even slightly superior to that of the EM method.