Comparison of iterative parametric and indirect deep learning‐based reconstruction methods in highly undersampled DCE‐MR Imaging of the breast

Greybox: A hybrid algorithm for direct estimation of tracer kinetic parameters from undersampled DCE-MRI data

BACKGROUND

A variety of deep learning-based and iterative approaches are available to predict Tracer Kinetic (TK) parameters from fully sampled or undersampled dynamic contrast-enhanced (DCE) MRI data. However, both the methods offer distinct benefits and drawbacks.

PURPOSE

To propose a hybrid algorithm (named as 'Greybox'), using both model- as well as DL-based, for solving a multi-parametric non-linear inverse problem of directly estimating TK parameters from undersampled DCE MRI data, which is invariant to undersampling rate.

METHODS

The proposed algorithm was inspired by plug-and-play algorithms used for solving linear inverse imaging problems. This technique was tested for its effectiveness in solving the nonlinear ill-posed inverse problem of generating 3D TK parameter maps from four-dimensional (4D; Spatial + Temporal) retrospectively undersampled k-space data. The algorithm learns a deep learning-based prior using UNET to estimate theK trans $\mathbf {K_{trans}}$ andV p $\mathbf {V_{p}}$ parameters based on the Patlak pharmacokinetic model, and this trained prior was utilized to estimate the TK parameter maps using an iterative gradient-based optimization scheme. Unlike the existing DL models, this network is invariant to the undersampling rate of the input data. The proposed method was compared with the total variation-based direct reconstruction technique on brain, breast, and prostate DCE-MRI datasets for various undersampling rates using the Radial Golden Angle (RGA) scheme. For the breast dataset, an indirect estimation using the Fast Composite Splitting algorithm was utilized for comparison. Undersampling rates of 8 × $\times$ , 12 × $\times$ and 20 × $\times$ were used for the experiments, and the results were compared using the PSNR and SSIM as metrics. For the breast dataset of 10 patients, data from four patients were utilized for training (1032 samples), two for validation (752 samples), and the entire volume of four patients for testing. Similarly, for the prostate dataset of 18 patients, 10 patients were utilized for training (720 samples), five for validation (216 samples), and the whole volume of three patients for testing. For the brain dataset of nineteen patients, ten patients were used for training (3152 samples), five for validation (1168 samples), and the whole volume of four patients for testing. Statistical tests were also conducted to assess the significance of the improvement in performance.

RESULTS

The experiments showed that the proposed Greybox performs significantly better than other direct reconstruction methods. The proposed algorithm improved the estimatedK trans $\mathbf {K_{trans}}$ andV p $\mathbf {V_{p}}$ in terms of the peak signal-to-noise ratio by up to 3 dB compared to other standard reconstruction methods.

CONCLUSION

The proposed hybrid reconstruction algorithm, Greybox, can provide state-of-the-art performance in solving the nonlinear inverse problem of DCE-MRI. This is also the first of its kind to utilize convolutional neural network-based encodings as part of the plug-and-play priors to improve the performance of the reconstruction algorithm.

Rapid 3D breath-hold MR cholangiopancreatography using deep learning-constrained compressed sensing reconstruction

OBJECTIVES

To compare the image quality of three-dimensional breath-hold magnetic resonance cholangiopancreatography with deep learning-based compressed sensing reconstruction (3D DL-CS-MRCP) to those of 3D breath-hold MRCP with compressed sensing (3D CS-MRCP), 3D breath-hold MRCP with gradient and spin-echo (3D GRASE-MRCP) and conventional 2D single-shot breath-hold MRCP (2D MRCP).

METHODS

In total, 102 consecutive patients who underwent MRCP at 3.0 T, including 2D MRCP, 3D GRASE-MRCP, 3D CS-MRCP, and 3D DL-CS-MRCP, were prospectively included. Two radiologists independently analyzed the overall image quality, background suppression, artifacts, and visualization of pancreaticobiliary ducts using a five-point scale. The signal-to-noise ratio (SNR) of the common bile duct (CBD), contrast-to-noise ratio (CNR) of the CBD and liver, and contrast ratio between the periductal tissue and CBD were measured. The Friedman test was performed to compare the four protocols.

RESULTS

3D DL-CS-MRCP resulted in improved SNR and CNR values compared with those in the other three protocols, and better contrast ratio compared with that in 3D CS-MRCP and 3D GRASE-MRCP (all, p < 0.05). Qualitative image analysis showed that 3D DL-CS-MRCP had better performance for second-level intrahepatic ducts and distal main pancreatic ducts compared with 3D CS-MRCP (all, p < 0.05). Compared with 2D MRCP, 3D DL-CS-MRCP demonstrated better performance for the second-order left intrahepatic duct but was inferior in assessing the main pancreatic duct (all, p < 0.05). Moreover, the image quality was significantly higher in 3D DL-CS-MRCP than in 3D GRASE-MRCP.

CONCLUSION

3D DL-CS-MRCP has superior performance compared with that of 3D CS-MRCP or 3D GRASE-MRCP. Deep learning reconstruction also provides a comparable image quality but with inferior main pancreatic duct compared with that revealed by 2D MRCP.

KEY POINTS

• 3D breath-hold MRCP with deep learning reconstruction (3D DL-CS-MRCP) demonstrated improved image quality compared with that of 3D MRCP with compressed sensing or GRASE. • Compared with 2D MRCP, 3D DL-CS-MRCP had superior performance in SNR and CNR, better visualization of the left second-level intrahepatic bile ducts, and comparable overall image quality, but an inferior main pancreatic duct.

VTDCE-Net: A time invariant deep neural network for direct estimation of pharmacokinetic parameters from undersampled DCE MRI data

PURPOSE

To propose a robust time and space invariant deep learning (DL) method to directly estimate the pharmacokinetic/tracer kinetic (PK/TK) parameters from undersampled dynamic contrast-enhanced (DCE) magnetic resonance imaging (MRI) data.

METHODS

DCE-MRI consists of 4D (3D-spatial + temporal) data and has been utilized to estimate 3D (spatial) tracer kinetic maps. Existing DL architecture for this task needs retraining for variation in temporal and/or spatial dimensions. This work proposes a DL algorithm that is invariant to training and testing in both temporal and spatial dimensions. The proposed network was based on a 2.5-dimensional Unet architecture, where the encoder consists of a 3D convolutional layer and the decoder consists of a 2D convolutional layer. The proposed VTDCE-Net was evaluated for solving the ill-posed inverse problem of directly estimating TK parameters from undersampled k - t $k-t$ space data of breast cancer patients, and the results were systematically compared with a total variation (TV) regularization based direct parameter estimation scheme. In the breast dataset, the training was performed on patients with 32 time samples, and testing was carried out on patients with 26 and 32 time samples. Translation of the proposed VTDCE-Net for brain dataset to show the generalizability was also carried out. Undersampling rates (R) of 8× , 12× , and 20× were utilized with PSNR and SSIM as the figures of merit in this evaluation. TK parameter maps estimated from fully sampled data were utilized as ground truth.

RESULTS

Experiments carried out in this work demonstrate that the proposed VTDCE-Net outperforms the TV scheme on both breast and brain datasets across all undersampling rates. For K trans $\mathbf {K_{trans}}$ and V p $\mathbf {V_{p}}$ maps, the improvement over TV is as high as 2 and 5 dB, respectively, using the proposed VTDCE-Net.

CONCLUSION

Temporal points invariant DL network that was proposed in this work to estimate the TK-parameters using DCE-MRI data has provided state-of-the-art performance compared to standard image reconstruction methods and is shown to work across all undersampling rates.

Radiomic Analysis of Pharmacokinetic Heterogeneity Within Tumor Based on the Unsupervised Decomposition of Dynamic Contrast-Enhanced MRI for Predicting Histological Characteristics of Breast Cancer

BACKGROUND

Breast tumor heterogeneity is associated with histological characteristics. However, pharmacokinetic (PK) heterogeneity within tumor might merit further exploration.

PURPOSE

To enhance the predictive power of molecular subtypes, Ki-67, and tumor grade by analyzing PK heterogeneity within tumor based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI).

STUDY TYPE

Retrospective.

POPULATION

Two hundred and eight biopsy-proven breast cancer patients, randomly divided into a training cohort (N = 144) and a testing cohort (N = 64).

FIELD STRENGTH/SEQUENCE

T₁ -weighted DCE-MRI at 3.0 T.

ASSESSMENT

A convex analysis of mixtures-compartmental modeling decomposition method was used to estimate the PK parameter (i.e., the volume transfer constant K^trans ) in tumor subregions with distinct physiological kinetic patterns, including fast-flow kinetics, slow-flow kinetics, and plasma input. Radiomic features based on the PK parameter were calculated from each tumor subregion.

STATISTICAL TESTS

The training cohort was used to build random forest classifiers based on the optimal features determined by the 5-fold cross-validation method. The performance was assessed on the testing cohort using the area under the receiver operating characteristic curve (AUC). The AUCs derived from the tumor subregion-based PK parameter were compared with those of the original images of the entire tumor using the DeLong test. A P-value of <0.05 was considered statistically significant.

RESULTS

The tumor subregion-based PK parameter, which yielded the highest AUCs of 0.8782, 0.7568, 0.7019, 0.7963, 0.8080, and 0.7375 for luminal A, luminal B, basal-like, human epidermal growth factor receptor 2, Ki-67, and tumor grade, respectively, obtained better diagnostic performance than the original images in the entire tumor (highest AUCs = 0.8612, 0.6191, 0.5593, 0.7704, 0.7494, and 0.6261, respectively). In particular, statistically significant improvement in the diagnostic performance was obtained for luminal B.

DATA CONCLUSION

Radiomic analysis of PK heterogeneity within tumor can enhance the predictive performance of radiomic models compared with that of the entire tumor.

LEVEL OF EVIDENCE

4 TECHNICAL EFFICACY STAGE: 3.

SpiNet: A deep neural network for Schatten p-norm regularized medical image reconstruction

PURPOSE

To propose a generic deep learning based medical image reconstruction model (named as SpiNet) that can enforce any Schatten p-norm regularization with 0 < p ≤ 2, where the p can be learnt (or fixed) based on the problem at hand.

METHODS

Model-based deep learning architecture for solving inverse problems consists of two parts, a deep learning based denoiser and an iterative data consistency solver. The former has either L2 norm or L1 norm enforced on it, which are convex and can be easily minimized. This work proposes a method to enforce any p norm on the noise prior where 0 < p ≤ 2. This is achieved by using Majorization-Minimization algorithm, which upper bounds the cost function with a convex function, thus can be easily minimized. The proposed SpiNet has the capability to work for a fixed p or it can learn p based on the data. The network was tested for solving the inverse problem of reconstructing magnetic resonance (MR) images from undersampled k space data and the results were compared with a popular model-based deep learning architecture MoDL which enforces L2 norm along with other compressive sensing-based algorithms. This comparison between MoDL and proposed SpiNet was performed for undersampling rates (R) of 2×, 4×, 6×, 8×, 12×, 16×, and 20×. Multiple figures of merit such as PSNR, SSIM, and NRMSE were utilized in this comparison. A two-tailed t test was performed for all undersampling rates and for all metrices for proving the superior performance of proposed SpiNet compared to MoDL. For training and testing, the same dataset that was utilized in MoDL implementation was deployed.

RESULTS

The results indicate that for all undersampling rates, the proposed SpiNet shows higher PSNR and SSIM and lower NRMSE than MoDL. However, for low undersampling rates of 2× and 4×, there is no significant difference in performance of proposed SpiNet and MoDL in terms of PSNR and NRMSE. This can be expected as the learnt p value is close to 2 (norm enforced by MoDL). For higher undersampling rates ≥6×, SpiNet significantly outperforms MoDL in all metrices with improvement as high as 4 dB in PSNR and 0.5 points in SSIM.

CONCLUSION

As deep learning based medical image reconstruction methods are gaining popularity, the proposed SpiNet provides a generic framework to incorporate Schatten p-norm regularization with 0 <p ≤ 2 with an added advantage of providing superior performance compared to its counterparts. The proposed SpiNet also has useful addition of Schatten p-norm value in regularization term being automatically chosen based on the available training data.