Synopsis

Keywords: Machine Learning/Artificial Intelligence, Brain, Hemoperfusion parameter estimation

Hemoperfusion magnetic resonance (MR) imaging derived parameters characterize both endothelial hyperplasia and neovascularization that are associated with tumor aggressiveness and growth. However, the hemoperfusion parameter estimation is still limited by low reliability, high bias, long processing time, and operator experience dependency up to now. In this study, a synthetic data driven learning method for hemoperfusion parameter estimation is proposed. Image analysis shows that the proposed method improves the reliability and precision of hemodynamic parameter estimation in a full-automatic and high-efficient way.

Introduction

Hemoperfusion magnetic resonance (MR) imaging techniques help quantify brain tumor microvessel proliferation and permeability, which is associated with tumor malignancy.¹ Dynamic susceptibility contrast magnetic resonance imaging (DSC-MRI) and dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) are two main hemoperfusion MR imaging techniques. Previous classical hemoperfusion parameter estimation methods suffer from limitations including large variance and bias in parameter predictions, long processing and computation time, and vulnerability to arterial input function (AIF) extraction operator experience.² Recent machine learning methods still depend on the large amount of real hemoperfusion MRI data and external algorithms for automatic AIF selection.³ Here we proposed a synthetic data driven learning method for full-automatic hemoperfusion parameter estimation that aims at addressing the currently existing hardships.

Methods

MR imaging: DSC- and DCE-MRI data of the human brain were acquired on a 3T SIEMENS Skyra scanner with a 16-channel head-neck coil. Detailed imaging acquisition protocol was as follows: (1) DSC-MRI: Single-shot gradient-echo EPI sequence, 220 × 220 mm² field of view, 5 mm slice sickness, 6 mm spacing between slices, TE/TR = 30/1600 ms, echo train length = 63, voxel size = 1.7 × 1.7 × 5.0 mm³. A total of 1,200 images were obtained in 96 seconds (20 sections, 60 phases). (2) DCE-MRI: 3D gradient-echo sequence, 220 × 220 mm² field of view, 5 mm slice sickness, 6 mm spacing between slices, TE/TR = 1.79/5.08 ms, voxel size = 0.7 × 0.7 × 5.0 mm³, 35 phases with temporal resolution = 5.9 s. Before the DCE-MRI, T₁ mapping images were acquired using the same pulse sequence and parameters except for varied flip angles = 2° and 15°.
Proposed scheme: Figure 1 illustrates the workflow of synthetic data generation for DSC- and DCE-MRI parameter estimation. The hemoperfusion parametric maps (CBV, CBF, K^trans and v_p) were first generated according to reasonable value distributions, then the time series of DSC- or DCE-MRI images were simulated according to the specific kinetic model of hemoperfusion. This enables the training data well-labeled and easily obtained, and facilitates the learning of nonlinear mapping between parametric maps and DSC/DCE image series for the deep neural network. It is worth noticing that AIFs for hemodynamic parameter estimation were automatically generated during data synthesizing, which facilitates the proposed method full-automatic.
Image analysis: Both DSC- and DCE-MRI have been proven reliable techniques in differentiating low-grade gliomas (LGGs) from high-grade gliomas (HGGs).⁴ To analyze the clinical diagnosis value of the proposed method, 29 patients with pathologically confirmed LGG or HGG were included in this study. DSC-MRI and DCE-MRI parameter estimations were also performed using the nordicICE software as reference (Nordic). Region of interest for glioma grading ability analysis was determined as the exact tumor parenchyma based on the contrast-enhanced T₁-weighted image, T₂-weighted and T₂-FLAIR images to avoid tumor necrosis, cystic areas and macro vessels. Receiver operating characteristic (ROC) analysis was employed to evaluate the diagnostic value. Theoretically, DSC-MRI derived CBV and DCE-MRI derived v_p provide similar physiological information.⁵ We assumed that there is a potential similarity between CBV and v_p, and calculated the structure similarity (SSIM) between them to confirm the estimation precision of the proposed method and Nordic. We deemed that the higher the SSIM is, the more precise the calculated parametric maps are.

Results

Figure 2 depicts the hemoperfusion parametric maps (CBV, CBF, K^trans and v_p) estimated from the proposed method and Nordic. All the hemoperfusion parametric maps estimated from the two methods are of high consistency in both LGG and HGG patients. Nonetheless, our proposed method improves the visual quality of all parametric maps with reduced noise and higher spatial coherency.
Figure 3 shows the ROC curves of hemodynamic parameters for differentiating HGG and LGG. The AUC values of CBV, CBF, v_p and K^trans estimated from the proposed method are 0.9483, 0.8988, 0.8920 and 0.8148, respectively, showing high diagnostic value. While for Nordic, the corresponding values are 0.7541, 0.7975, 0.8302 and 0.7747, which is apparently lower than the proposed method.
Figure 4 illustrates CBV and v_p maps of 6 HGG patients and 5 LGG patients estimated from the proposed method and Nordic, together with the corresponding SSIM. For the proposed method, the estimated v_p map is visually more similar to the CBV map than Nordic for both LGG and HGG patients. Moreover, the proposed method achieves higher SSIM between the DSC-MRI derived CBV and DCE-MRI derived v_p for all 11 subjects compared to Nordic, implying higher hemodynamic parameter estimation precision.

Discussion and conclusion

In the present study, we put forward a full-automatic method for DSC- and DCE-MRI parameter estimation with synthetic data. Apart from the better parametric map quality, the performance in glioma grading of the proposed method shows a noticeable margin over the widely acknowledged software Nordic with obviously higher AUC of ROC curves. The higher SSIM between the proposed method estimated CBV and v_p compared to Nordic further confirms a more precise estimation. Moreover, the proposed method reduces the estimation time from several minutes to tens of milliseconds. In summary, the proposed method improves the reliability and precision of hemodynamic parameter estimation in a full-automatic and high-efficient way.

Acknowledgements

This work was supported by the National Natural Science Foundation of China under grant numbers 11775184, 82071913, and 22161142024.