A Novel Self-Supervised Learning-Based Method for Dynamic CT Brain Perfusion Imaging

Abstract

Dynamic computed tomography (CT)-based brain perfusion imaging is a non-invasive technique that can provide quantitative measurements of cerebral blood flow (CBF), cerebral blood volume (CBV), and mean transit time (MTT). However, due to high radiation dose, dynamic CT scan with a low tube voltage and current protocol is commonly used. Because of this reason, the increased noise degrades the quality and reliability of perfusion maps. In this study, we aim to propose and investigate the feasibility of utilizing a convolutional neural network and a bi-directional long short-term memory model with an attention mechanism to self-supervisedly yield the impulse residue function (IRF) from dynamic CT images. Then, the predicted IRF can be used to compute the perfusion parameters. We evaluated the performance of the proposed method using both simulated and real brain perfusion data and compared the results with those obtained from two existing methods: singular value decomposition and tensor total-variation. The simulation results showed that the overall performance of parameter estimation obtained from the proposed method was superior to that obtained from the other two methods. The experimental results showed that the perfusion maps calculated from the three studied methods were visually similar, but small and significant differences in perfusion parameters between the proposed method and the other two methods were found. We also observed that there were several low-CBF and low-CBV lesions (i.e., suspected infarct core) found by all comparing methods, but only the proposed method revealed longer MTT. The proposed method has the potential to self-supervisedly yield reliable perfusion maps from dynamic CT images.

Keywords: Bi-directional long short-term memory; Computed tomography perfusion; Convolutional neural network; Self-supervised learning.

Fig. 1

The architecture of the proposed CNN-BiLSTM model. The custom layer that implements Eq. 6 is responsible for calculating the predicted TAC from the measured AIF(t) and the predicted $\tilde{K}\left(t\right)$

Fig. 2

Perfusion maps of CBF, CBV, and MTT (columns) obtained using SVD, TTV, and CNN-BiLSTM methods (second, third, and fourth rows), compared with the ground truth (first row), for the simulated CT perfusion data (I = 95mA)

Fig. 3

MARE values of CBF (top row), CBV (middle row), and MTT (bottom row) for each studied method and each tissue type at three different simulated noise levels

Fig. 4

Comparison of the simulated (true) IRF curves with the predicted IRF curves obtained from the three studied methods for GM, WM, and infarct core for the simulated CT data with I = 47.5 mA

Fig. 5

Perfusion maps of CBF, CBV, and MTT (columns) of one patient obtained using SVD (top rows), TTV (middle rows), and CNN-BiLSTM (bottom rows) methods

Fig. 6

Perfusion maps of CBF, CBV, and MTT (columns) of one patient obtained using SVD (top rows), TTV (middle rows), and CNN-BiLSTM (bottom rows) methods

Fig. 7

Perfusion maps of CBF, CBV, and MTT (columns) of one patient obtained using SVD (top rows), TTV (middle rows), and CNN-BiLSTM (bottom rows) methods

Fig. 8

Perfusion maps of CBF, CBV, and MTT (columns) of one patient obtained using SVD (top rows), TTV (middle rows), and CNN-BiLSTM (bottom rows) methods

Fig. 9

Perfusion maps of CBF, CBV, and MTT (columns) of one patient obtained using SVD (top rows), TTV (middle rows), and CNN-BiLSTM (bottom rows) methods

Fig. 10

Perfusion maps of CBF, CBV, and MTT (columns) of one patient obtained using SVD (top rows), TTV (middle rows), and CNN-BiLSTM (bottom rows) methods

Fig. 11

Perfusion maps of CBF, CBV, and MTT of one patient from the ISLES 2018 training data obtained using the SVD, TTV, and CNN-BiLSTM methods and the contour of the expert segmentation of the infarct lesion in magenta

Fig. 12

Perfusion maps of CBF, CBV, and MTT of one patient from the ISLES 2018 training data obtained using the SVD, TTV, and CNN-BiLSTM methods and the contour of the expert segmentation of the infarct lesion in magenta

Fig. 13

Perfusion maps of CBF, CBV, and MTT of one patient from the ISLES 2018 training data obtained using the SVD, TTV, and CNN-BiLSTM methods and the contour of the expert segmentation of the infarct lesion in magenta

Fig. 14

Perfusion maps of CBF, CBV, and MTT (columns) of three patients (corresponding to Figs. 4, 5, and 6) obtained using the SPPINN method

Fig. 15

Perfusion maps of CBF, CBV, and MTT (columns) of three patients (corresponding to Figs. 7, 8, and 9) obtained using the SPPINN method