Purpose

Pharmacokinetic parameters from quantitative DCE-MRI provide characterization of the tissue perfusion and flow, and are a potential biomarker for tumor staging and prognosis^1,2. However, most standard-of-care DCE-MRI protocols have a limited number of phases and insufficient temporal resolution for quantitative analysis³. Advanced quantitative DCE-MRI techniques (e.g. Multitasking DCE⁴, QIN⁵, etc.) usually require specialized sequences and lack sufficient clinical validation for wide clinical application. The objective of this study is to develop a retrospective quantification method for multi-phasic DCE-MRI by deep learning. In this study, we synthesized clinical multi-phasic breast DCE-MRI by downsampling 2.8-s temporal resolution Multitasking DCE and developed a deep learning-based method for retrospective pharmacokinetic quantification. We evaluated method performance in tumor mass and contralateral non-tumor breast tissue in the Duke-Breast-Cancer-MRI dataset⁶.

Materials and Methods

Data description:
The datasets for training and validation, 19 female subjects, including 13 healthy volunteers and 6 patients with pathologically confirmed triple-negative breast cancer, were scanned on a 3T scanner (MAGNETOM Vida, Siemens Healthcare, Germany). A 20% low-dose (0.02mmol/kg) Multitasking DCE protocol was first performed. Afterwards, a 20% low-dose repeat study was performed on 10 of the healthy volunteers, and a standard-dose (0.1 mmol/kg) Multitasking DCE protocol was performed on 9 of the healthy volunteers. Low-dose or standard-dose Multitasking DCE acquisition lasted for 10-min with 0.9 mm x 0.9 mm x 1.1 mm spatial resolution. Dynamic T1 maps with 2.8-s temporal resolution were obtained from Multitasking DCE.
Figure 1 (A) summarizes the synthesis of the input and reference. Each T1W DCE-MRI signal $$$S_{ref}$$$ was generated from 2.8-s temporal resolution both doses of Multitasking R1 values using the FLASH signal equation, then downsampled to multi-phasic DCE-MRI $$$S_{in}$$$ based on a randomized number of phases (4 to 7) and temporal resolution (74-s to 130-s). The same pipeline was used for left ventricular so that 2.8-s temporal resolution $$$AIF_{ref}$$$ and multi-phasic artery input function $$$AIF_{in}$$$ were generated. The synthesis was based on the median TR/TE/flip angle in the Duke-Breast-Cancer-MRI dataset (TR/TE=5.2/2.39 ms, flip angle=10°).
The standard Tofts model⁷ was used. A pre-screening strategy was used to exclude data that could not be well characterized by the standard Tofts model, i.e. pixel-wise data with an L-S fitting residual of greater than 1 was excluded.
Network architecture:
A deep-learning-based pipeline was built to enable retrospective quantification by improving the temporal resolution. The neural network consisted of two sub-networks. As shown in Figure 1(B), upsampling block aims to convert low-temporal resolution inputs to high-temporal resolution outputs, which were passed to a deconvolution block along with $$$AIF_{ref}$$$ for three pharmacokinetic parameters outputs $$$K^{trans}$$$, $$$k_{ep}$$$, and $$$v_e$$$ in the standard Tofts model.
The whole network was trained jointly in 100 epochs. The loss function consists of three components: (1) mean squared error; (2) data consistency loss defined as mean squared error of outputs and reference in known phases; (3) self-supervised loss defined as mean squared error of outputs and reconstructed outputs using the intermediate pharmacokinetic parameters and $$$AIF_{ref}$$$. In brief, the cost function was:
$$$L=\left\|S_{\text {out }}-S_{\text {ref }}\right\|_2^2+\lambda_1\left\|S_{\text {in }}-\Omega S_{\text {out }}\right\|_2^2+\lambda_2\left\|S_{\text {out }}-f_{\text {Tofts }}\left(A I F_{\text {out }}, k_{\mathrm{ep}, \text { out }}, K_{\text {out }}^{\text {trans }}\right)\right\|_2^2$$$
with standard Tofts model $$$f_{Tofts}\left(C_p, k_{ep}, K^{trans}\right)=K^{trans} C_{\mathrm{p}} * e^{-k_{\mathrm{ep}} t}$$$ and temporal down-sampling operator $$$\Omega$$$.
For inference, only the upsampling block was used for temporal super-resolution before parameter estimation by variable projection (VARPRO)⁸. Tissue signal $$$S_{out}$$$ and the $$$AIF_{out}$$$ were obtained by passing $$$S_{in}$$$ and $$$AIF_{in}$$$ to the upsampling block individually.
Performance assessment:
The Duke-Breast-Cancer-MRI dataset was used to assess the performance of the proposed method. An experienced radiologist labeled tumor and non-tumor regions in 89 of the dataset. Testing was performed in these labeled patient data.

Result

A representative example of retrospectively quantified pharmacokinetic maps from one of the breast cancer patients in the test set was shown in Figure 2. The tumor mass is marked by a red arrow, and the non-tumor tissue is marked by a green arrow. Retrospectively quantified maps demonstrated higher $$$K^{trans}$$$, higher $$$k_{ep}$$$, and higher $$$v_e$$$ in tumor region compared to contralateral tissue. Figure 3 shows the box plot of the tumor region versus non-tumor region. Table 1 shows the ROI statistics of the retrospective quantified parameters in tumor and non-tumor regions. Both figure and table show that the proposed method provides promising results in terms of uniformity within the same tissue and difference between different tissue types. For all three parameters, Wilcoxon Signed Rank test shows significant differences.

Discussion and Conclusion

In this work, we demonstrated the feasibility of retrospective quantification of multi-phasic clinical breast DCE-MRI. The quantitative pharmacokinetic parameters derived retrospectively using deep learning show significant differences between tumor and contralateral tissue, which is consistent with previous studies⁹ using quantitative DCE MRI. This retrospective approach has the potential to quantify a large number of conventional multi-phasic breast DCE MRI studies to establish the range of various pharmacokinetic parameters for tumor and non-tumor tissues. Upon further validation, this approach could quantify the underlying pharmacokinetics using clinical multi-phasic DCE MRI to improve breast cancer diagnosis and assessment of treatment responses.

Acknowledgements

No acknowledgement found.