Purpose

Pharmacokinetic parameters from quantitative DCE-MRI are potential biomarkers to help differentiate tumor stages and assess prognosis before treatment^1,2. However, clinical multi-phasic abdominal DCE-MRI has limited temporal resolution (approximately 30-sec) and number of samples (n=4), and thus can only provide qualitative or semi-quantitative assessments³. The long-term objective of this project is to develop a retrospective quantification method for multi-phasic abdominal DCE-MRI by improving the temporal resolution via deep learning. In this proof-of-concept study, we will simulate conventional multi-phasic abdominal DCE-MRI by downsampling 2-sec temporal resolution Multitasking DCE⁴ data and assess the accuracy of retrospectively estimated kinetic parameters against the ground truth.

Materials and Methods

Inputs and labels preparation
40 subjects (16 healthy volunteers, 14 patients with pancreatic ductal adenocarcinoma (PDAC), 10 patients with pancreatitis) were scanned on a 3T scanner (MAGNETOM Vida, Siemens Healthcare, Germany). Six-dimensional free-breathing Multitasking DCE^4,5 (MT-DCE) was used to obtain dynamic T1 mapping at 2-sec temporal resolution. The total imaging time was 12 min; 0.1 mmol/kg contrast media was administered 2-min into the scan. Other imaging parameters include: spatial resolution=1.4x1.4x3.0 mm³, number of slices=96.
Figure 1(A) summarizes the simulation of conventional multi-phasic DCE-MRI from Multitasking DCE. Multi-phasic DCE-MRI is based on the following clinical protocol⁶: 3D breath-holding T1W GRE, temporal resolution 30 s (TA 10 s, rest time 20 s), TR 2.8 ms, TE 1.1 ms, Cartesian linear sampling, one pre-contrast phase, and three post-contrast phases. The simulated T1W signals were generated by the GRE signal equation and served as inputs to the deep learning model. 2-sec temporal-resolution T1W signals simulated from the Multitasking data served as labels during training. To convert T1W signal intensity to contrast agent concentration for parametric quantification, both input and label signals were baseline-corrected and normalized by the pre-contrast signal $$$S_0$$$. A total of ~57,891 voxels were used for training and validation.
Network architecture and training
A deep neural network was constructed for temporal super-resolution and kinetic mapping. As shown in Figure 1(B), 30-s temporal-resolution $$$S(t)$$$ was passed to an upsampling block and compared with the 2-s temporal-resolution label $$$S(t)$$$. Data consistency loss was added to enforce the consistency between the inputs and the corresponding samples of the labels. For quantification, the temporally-super-resolved $$$S(t)$$$ and arterial input function (AIF) was passed to a deconvolution block to output $$$k_{ep}$$$ and $$$K^{trans}$$$ was then calculated as $$$K^{trans}\ /k_{ep}$$$. This deconvolution block was trained entirely self-supervised without quantitative labels, such that the mapped $$$k_{ep}$$$ and $$$K^{trans}$$$ were consistent with the super-resolution signal output. In brief, the cost function was:$$L=\left\|S_{\mathrm{out}}-S_{\mathrm{label}}\right\|_{2}^{2}+\lambda_{1}\left\|S_{\mathrm{in}}-\Omega S_{\mathrm{out}}\right\|_{2}^{2}+\lambda_{2}\left\|S_{\text {out }}-f_{\mathrm{Tofts}}\left(A I F_{\text {out }}, k_{\mathrm{ep}, \mathrm{out}}, K_{\mathrm{out}}^{\mathrm{trans}}\right)\right\|_{2}^{2}$$ with Tofts model⁷ $$$f_{Tofts}(C_p,k_{ep},K^{trans})=K^{trans}C_p*e^{-k_{ep}t}$$$ and temporal down-sampling operator $$$\Omega$$$.
The network was trained with Adam optimizer⁸ with an initial learning rate of 10^-3. The learning rate was reduced by a factor of 10 when the validation loss stopped improving in 10 continuous epochs. The training was ended in 100 epochs.
Testing
For testing, pharmacokinetic parameters were fitted from Multitasking images using least-square fitting of the Tofts model with variable projection (VARPRO⁹). These parameters served as testing labels.

Results

Representative examples
A representative example of temporal super-resolution from a 57-year-old subject with PDAC was shown in Figure 2(A). The tumor mass was marked by a yellow boundary on the simulated artery phase image and the abdominal aorta was depicted by a red dashed circle. One example of retrospectively quantified and ground truth pharmacokinetic maps from the same subject was shown in Figure 2(B). The overlaid colored voxels show the pharmacokinetic maps from retrospective quantification (top) and Multitasking DCE data (bottom). The yellow arrow points out a benign cyst while the red arrow points out the tumor mass.
Agreement between estimated parameters and ground truth
The correlation between retrospectively quantified kinetic maps from the proposed method and the ground truth from Multitasking DCE is shown in Figure 3A, with Bland-Altman plots shown in Figure 3B. The intraclass correlation coefficient (ICC), the coefficient of determination (R), and the coefficient of variation (CoV) were also evaluated.
Ability to differentiate abnormal tissue
Figure 4 contains a box plot that shows the median and quartiles of the retrospectively quantified parameters (left) and ground truth (right) from the control, tumor, nontumoral and pancreatitis regions. Paired t-tests were used to evaluate the mean difference between different regions. Tumor and non-tumor regions showed significant differences in $$$v_e$$$, $$$K^{trans}$$$, $$$k_{ep}$$$ estimated by the proposed method (P= <0.01, <0.01, <0.01, respectively). The control region showed significant differences comparing to tumor region in $$$K^{trans}$$$ and $$$k_{ep}$$$ (P = <0.01, <0.01, respectively), and to pancreatitis region in $$$v_e$$$, $$$K^{trans}$$$, $$$k_{ep}$$$ (P= <0.01, <0.01, <0.01, respectively). Retrospectively quantified $$$K^{trans}$$$, $$$k_{ep}$$$ showed significant difference in control and non-tumor regions (P= <0.01, 0.043 respectively). $$$v_e$$$ estimated by the proposed method showed significant difference in control and non-tumor regions (P=0.03) while $$$v_e$$$ of ground truth did not (P=0.08). These findings were all consistent with the ground truth and were in general agreement with published findings^10,11.

Conclusion and Discussion

We demonstrated that quantification of clinical multiphasic DCE-MRI is possible by deep learning. Future work will prospectively acquire MT-DCE and clinical multiphasic DCE images on the same subjects and validate the approach. If validated, this technique has the potential to improve cancer diagnosis.

Acknowledgements

This work was supported in part by the National Institutes of Health (Grant/Award Nos. R01EB028146 and R01HL156818).