Proper spatial alignment of anatomical landmarks during and between liver imaging exams is challenging due to the dynamic morphology of the liver. Liver-focused registration algorithms have been developed but are typically semiautomatic. We propose a fully-automated pipeline for affine-based registration of inter- and intra-exam liver images and assess performance on clinical liver MRI exams at 1.5T and 3T. The proposed pipeline achieved comparable or superior accuracy and scalability to that reported for previously proposed algorithms. Retrospective image review by an expert abdominal radiologist confirmed subjective improvement in anatomic registration and lesion co-localization. Proof of concept of multimodal scalability was demonstrated.
Introduction
Medical image registration (MIR) increases radiologist confidence and diagnostic accuracy1. However, proper spatial alignment of liver and liver lesions during and between imaging exams is challenged by the dynamic morphology of the liver and its surrounding organs as well as unavoidable variability in patient positioning, movement, and physiological motion2-4.
Both affine and deformable liver-focused MIR algorithms have been developed to address these challenges but are typically semiautomatic and slow due to CPU implementation5-9. Moreover, the majority of liver-focused MIR algorithms were tested on small numbers of patients and may not be generalizable. Convolutional neural networks (CNN) have proven fast and effective for deformable MIR10-11, but deformable methods are unreliable due to their exceedingly flexible parameterizations12-14.
We propose a fast two-step, fully automated pipeline for affine-based registration of inter- and intra-exam liver images comprising a liver segmentation CNN and affine transformer (AT) parallelized on a GPU for speed15. Performance is evaluated on a large cohort of clinical liver MRI exams through a variety of image similarity metrics. Clinical utility is illustrated through case examples. Multimodal scalability is explored.
Methods
Imaging Data
This retrospective, single-center, HIPAA-compliant, IRB-approved study included 1016 image acquisitions from 253 unique patients who underwent a total of 564 gadoxetate-enhanced 1.5T or 3T liver MRI exams from 2011 to 2018. The dataset comprised baseline and follow-up 3D T1-weighted hepatobiliary phase (HBP) acquisitions, including multiple intra-exam acquisitions. All possible within-patient pairs of acquisitions, including those acquired in the same exam, were used for registration validation, producing 2225 acquisition pairs.
Registration Algorithm
Acquisition pairs, comprising a static acquisition (e.g., baseline acquisition) and a moving acquisition (e.g., follow-up acquisition), are first sent to an independently developed liver segmentation CNN with U-net architecture to create intensity-populated liver masks (intensity-masks), which are used to constrain the registration to focus on the liver (Figure 1). Intensity-masks are registered using an AT network with a single 12-neuron dense layer representing affine parameters. Estimated parameters are used to map the native moving acquisition to static acquisition space. Since the registration pipeline is individually applied to each acquisition pair and is fully unsupervised, we do not require a leave-out dataset for testing.
Evaluation
Binary liver masks, intensity-masks, and whole images were used as input into the proposed pipeline to study the impact of liver-focused inputs on registration accuracy. Performance was evaluated by computation time, image similarity metrics, and subjectively by an expert radiologist through retrospective image review. Differences in image similarity metrics were statistically evaluated using Wilcoxon signed-rank tests due to departures from normality.
Results
Median [10th, 90th percentile] intersection-over-union (IoU) percentages were 87.35% [70.71%, 94.08%] for binary-mask, 86.73% [69.37%, 94.05%] for intensity-mask, and 77.90% [38.08%, 92.23%] for whole-image inputs. Intensity correlations (ICs) were 0.91, [0.79, 0.97] for binary-mask, 0.91 [0.80, 0.97] for intensity-mask, and 0.85, [0.52, 0.96] for whole-image inputs. Liver-focused inputs produced significantly higher median IoU and IC than whole-image inputs. Binary-mask inputs provided significantly higher IoU and lower IC than intensity-mask inputs, but differences were small. Binary-mask inputs achieved convergence (49.75s [40.99s, 67.18s]) significantly faster than intensity-mask (84.56s [66.68s, 119.98s]) or whole-image (75.51s [55.05s, 112.86s]) inputs. Subjectively, manual registration by an expert radiologist resulted in ambiguous lesion localization and an inability to register all parts of the liver simultaneously. The affine-registration mitigated these problems (Figures 2-4) and improved radiologist confidence for lesion localization.Discussion
Liver-focused MIR constrains the algorithm to maximize spatial alignment across the liver region only, producing more accurate registration than whole-image inputs. Minor differences in binary-mask and intensity-mask performance suggest liver-masking drives the registration, with inclusion of intensities providing only marginal improvement. An accurate fully-automated, liver-focused pipeline for CT/open-MR registration has been proposed6. However, performance was validated across only 18 livers and produced an average liver overlap of 78%, with prolonged runtime (540s). Our study included 2225 acquisition pairs from 253 patients and produced a comparable overlap score (87%), with median runtime <60s for binary-mask input. Preliminary application across different imaging phases and modalities shows proof-of-concept of multimodal scalability (Figure 5). Additionally, affine-registered images improved reader confidence, which should be further confirmed in prospective studies.Conclusion
Our proposed liver registration pipeline achieves comparable or superior accuracy and scalability to previously proposed algorithms. Fully-automated liver registration can automate intra-exam lesion characterization and improve reader confidence of inter-exam comparisons16-18. Prospective research is necessary to assess the large-scale impact of registered liver images on patient outcomes when incorporated into clinical radiology workflow.Figure 2a-b: Baseline and follow-up: Slices are manually registered using the bifurcation of the right portal vein as anatomical reference (yellow arrows). Note difference in morphology of posterior aspect and left lobe of liver (orange circles) due to different liver position.
Figure 2c: Registered follow-up: Posterior aspect and left lobe of liver are more consistent with appearance on baseline image.
Figure 4a-b: Baseline and follow-up: Images manually registered using anatomical landmarks (yellow arrows). Lesion (white arrow) observed in follow-up image without correspondence on baseline scan.
Figure 4c: Affine registered image shows correspondence of the anatomical landmarks.
Figure 4d-e: Baseline and follow-up images registered using left portal vein as anatomical landmark. Nodule seen on baseline scan is not shown in aligned follow-up slice. Confidence on lesion correspondence is low due to differences in lesion size and liver appearance.
Figure 4f: Following image registration, the nodular image can be confidently confirmed as a preexisting lesion that grew from baseline.
Figure 5a-c: Cross-modality image registration. T1 weighted MR Image (a) is registered to previous CT study (b). Note difference in the morphology of the posterior aspect and left lobe of the liver, which is corrected using the proposed registration (c).
Figure 5d-f: Intra-exam image registration. Hepatobiliary phase image (d) is registered to arterial phase image (e). Note difference in the morphology of the left lobe of the liver and the presence of a third lesion not seen in the corresponding arterial phase slice. After affine registration anatomical landmarks and liver metastasis are colocalized.