Performance of U-net based pyramidal lucas-kanade registration on free-breathing multi-b-value diffusion MRI of the kidney

Abstract

Objective: In free-breathing multi-b-value diffusion-weighted imaging (DWI), a series of images typically requires several minutes to collect. During respiration the kidney is routinely displaced and may also undergo deformation. These respiratory motion effects generate artifacts and these are the main sources of error in the quantification of intravoxel incoherent motion (IVIM) derived parameters. This work proposes a fully automated framework that combines a kidney segmentation to improve the registration accuracy.

Methods: 10 healthy subjects were recruited to participate in this experiment. For the segmentation, U-net was adopted to acquire the kidney's contour. The segmented kidney then served as a region of interest (ROI) for the registration method, known as pyramidal Lucas-Kanade. Our proposed framework confines the kidney's solution range, thus increasing the pyramidal Lucas-Kanade's accuracy. To demonstrate the feasibility of our presented framework, eight regions of interest were selected in the cortex and medulla, and data stability was estimated by comparing the normalized root-mean-square error (NRMSE) values of the fitted data from the bi-exponential intravoxel incoherent motion model pre- and post- registration.

Results: The results show that the NRMSE was significantly lower after registration both in the cortex (p < 0.05) and medulla (p < 0.01) during free-breathing measurements. In addition, expert visual scoring of the derived apparent diffusion coefficient (ADC), f, D and D* maps indicated there were significant improvements in the alignment of the kidney in the post-registered image.

Conclusion: The proposed framework can effectively reduce the motion artifacts of misaligned multi-b-value DWIs and the inaccuracies of the ADC, f, D and D* estimations. Advances in knowledge: This study demonstrates the feasibility of our proposed fully automated framework combining U-net based segmentation and pyramidal Lucas-Kanade registration method for improving the alignment of multi-b-value diffusion-weighted MRIs and reducing the inaccuracy of parameter estimation during free-breathing.

Figure 1.

Pipeline of our proposed framework. Multi-b-value DWIs were segmented by U-net and then registered by pyramidal Lucas-Kanade registration method. DWI, diffusion-weighted imaging.

Figure 2.

Eight ROIs positioned on anatomical image (b = 0). The red and yellow circles indicate parts of the cortex and medulla, respectively. ROIs, regions of interest.

Figure 3.

(a) Training and validation dice values for each training and validation epoch. (b) Segmentation result (red edge) with manual ground truth (green border).

Figure 4.

Registration of the target image to the reference image of two typical subjects. (a) Subject #1. Left: Reference image (frame 1, b = 0 s mm^–2). For comparison, the contour of the kidney (green) from frame 1 is copied in its original location in each subsequent frame. Pre-(middle) and Post-(right) registration images of the same slice from frame (b = 50 s mm^–2). (b) Subject #2. Left: Reference image (frame 1, b = 0 s mm^–2). Similarly, the contour of the kidney (green) from frame 1 is copied in its original location in each subsequent frame. Pre- (middle) and Post-(right) registration images of the same slice from frame (b = 550 s mm^–2).

Figure 5.

Computed f, D, D* and ADC maps of two typical subjects (the arrows indicate some major differences). In contrast to the non-processed images, the visual quality of the four computed parameter maps was clearly improved in the registration case. ADC, apparent diffusion coefficient.

Figure 6.

Comparison of the NRMSE of the cortex and medulla between the original and registered measurements for the eight ROIs of ten subjects. NRMSE, normalized root-mean-square error; ROIs, regions of interest.