Skip to main content
2025 SCMR Conference Main banner
Final Program


Schedule at a Glance
Back

Rapid Fire Abstracts

Rapid Fire 01- Image Reconstruction I

Elimination of Motion-induced Subendocardial Dark-rim Artifacts in Stress Perfusion CMR Enabled by Spatiotemporal Deep Learning (RF_TH_158)

Thursday, January 30, 2025
11:10 AM – 11:20 AM East Coast USA Time
Location: Blue Room PreFunction
Figure 1: Proposed technique for elimination of motion-induced subendocardial dark-rim artifacts. Step 1: the undersampled k-space from the perfusion image series is reconstructed to fully encoded k-space using the default scanner recon method, e.g., conventional GRAPPA with partial Fourier reconstruction. Step 2: The reconstructed fully sampled k-space is then used to generate two different recons by applying a temporal footprint (TF) reduction technique by removing consecutive phase-encode lines from one (left or right) side of the k-space and applying partial Fourier recon. Step 3: a spatiotemporal deep neural network (2D+time UNet-based architecture) is used to rapidly segment the myocardium into 12 sectors (standard AHA segmentation with endo-epi subsegments) and extract the average signal intensity (normalized by peak blood pool signal) for both left-sided and right-sided recons (shown as the top and bottom gray-scale bull’s eye maps, respectively). Next, the slope of the line connecting the left- vs right-sided S.I. for each segment is computed to generate a bull’s eye map of dark-rim TF variability (color-coded bull’s eye map). To automatically decide the optimal TF truncation (i.e., left vs right), the dark-rim TF variability is thresholded and combined with a “majority voting” approach (see the bull’s eye map with “L” and “R” votes) to automatically decide the optimal TF truncation (i.e., to choose the left vs right recon solution). The result is a DRA-suppressed images series (independently generated for each slice position) with minimal motion-induced DRA for each slice position.

Figure 2: Representative results for the proposed DRA elimination method. (A) These are two cases from the 10 stress FPP studies selected retrospectively based on “truly normal” invasive testing using a comprehensive protocol in the catheterization laboratory including: normal quantitative coronary angiography (no epicardial coronary stenosis > 20%), normal invasive coronary function testing (normal adenosine coronary flow reserve and normal coronary endothelial function), normal LV filling pressures and a negative LGE. The key feature of this validation approach is that it enables a notably more objective approach for DRA scoring since the stress scans are expected to be completely normal. As can be seen, in both cases, the scanner default recon has noticeable subendocardial DRA that can affect the assessment given the thin transmural extent (diastolic frame); however, the automatically generated optimal recon (green boxes) effectively eliminates the motion-induced DRA by selecting the optimal TF truncation approach (right-sided for Case 1 and left-sided for Case 2). In contrast, the non-optimal side results in a severe DRA (red arrows), which shows the impact of the TF choice (left vs right selection) on the presence and severity of DRA due to motion-induced (point spread function) effects. (B): Two examples from the feasibility testing of the inline implementation (FIRE framework) of the proposed technique. Both examples are clinically-indicated stress FPP studies on a 1.5T scanner (MAGNETOM Sola, Siemens Healthineers AG, Forchheim, Germany) wherein an additional image series is generated (besides the scanner default recon) with automatically generated DRA-optimal recon.

Figure 3: Summary of comparison between the scanner default recon and the proposed DRA-optimal recon in terms of (A) DRA severity, and (B) prevalence of severe DRA. By leveraging the availability of comprehensive invasive coronary angiography and coronary function testing in the selected subjects (n=10), the image artifact scores were assigned (DRA severity on 0-4 scale; 0: no artifact) by two expert readers in consensus. (A) DRA severity was significantly lower for the DRA-optimal approach vs default recon (1.5 ± 0.72 vs 2.2 ± 0.87, p <0.005) whereas the DRA severity for the non-optimal recon (i.e., selecting the opposite left/right side vs the optimal choice) is significantly higher than both (3.3 ± 0.74, p<0.001). (B) The percentage of slices with severe DRA (artifact score >2) was 47% for conventional method, 83% for non-optimal reconstruction and 7% for proposed DRA-optimal approach (p < 0.01 for all pairs. The proposed DRA-optimal reconstruction coupled with inline implementation provides an easily deployable DRA-reduction technique that can be applied to any standard- or high-resolution Cartesian-sampled FPP pulse sequence.


Background: The subendocardial dark-rim artifact (DRA) remains an important challenge in routine clinical use of stress first-pass perfusion (FPP) CMR as it reduces the diagnostic accuracy in patients with suspected ischemia by mimicking the appearance of subtle perfusion defects. DRA is primarily induced by Gibbs ringing1 and cardiac motion.2 Several methods have been proposed to suppress Gibbs-ringing-induced DRA3 and to use optimized or non-Cartesian k-space sampling.4-6 However, a method to optimize the image reconstruction (“recon”) pipeline to automatically suppress motion-induced DRAs for routine Cartesian FPP studies is lacking. We propose a deep learning (DL)-enabled method with validation against invasive testing and inline implementation.

Methods: Fig. 1 describes the proposed method. In Step 1, accelerated stress FPP data is reconstructed (e.g., using the default scanner recon routine). Step 2 performs two different recons by first applying a one-sided (left or right) temporal footprint (TF) reduction (Fig. 1) to modify the effective TF of the acquired data followed by partial Fourier recon. In Step 3, a spatiotemporal DL network is employed to segment the free-breathing time frames and to extract the segment-wise signal intensity (S.I.) for both the left- and right-sided recons. Next, the slope of the line connecting the left- vs right-sided S.I. for each segment is computed to generate a bull’s eye map of dark-rim TF variability.7 Finally, the TF variability is thresholded and combined with a “majority voting” approach to automatically decide the optimal TF truncation (i.e., left vs right) for recon of DRA-suppressed images series. The method was validated on a retrospective cohort of patients with normal coronary angiograms (no stenosis >20% diameter) and normal invasive coronary function testing (normal microvascular reactivity including endothelial function and flow reserve > 2.5) and negative LGE (n=10 stress studies). In this “truly normal” cohort, the DRA severity was evaluated on a visual 0-to-4 scale (0: no artifact). Further, the feasibility of inline implementation (Siemens FIRE framework)8 was prospectively tested (n=8 stress studies).

Results: Fig. 2A shows representative results for two of the truly normal patient studies. In both cases, the scanner default recon has noticeable DRA that can affect the assessment given the thin transmural extent (diastolic frame); however, the automatically generated optimal recon (green boxes) effectively eliminates the motion-induced DRA by selecting the optimal TF truncation approach (right-sided for Case 1 and left-sided for Case 2). In contrast, the non-optimal side results in a severe DRA (red arrows), which shows the impact of the TF choice (left vs right) on DRA. Fig. 2B shows inline implementation examples demonstrating feasibility in clinical routine. Fig. 3 summarizes overall DRA scores for default vs proposed approach showing significant improvement in DRA severity (2.2 ± 0.87 vs 1.5 ± 0.72, p < 0.005) and DRA prevalence (47% vs 7%, p < 0.01).


Conclusion:

The proposed DL-enabled recon method for stress FPP studies results in significant reduction in DRA severity based on a rigorously designed evaluation (truly normal cases). Coupled with inline implementation, this technique may provide an easily deployable DRA-reduction technique in routine clinical studies with standard- or high-resolution CMR protocols.9,10 To the best of our knowledge, this is the first inline method for suppression/elimination of DRAs in stress FPP studies.

Figure 1: Proposed technique for elimination of motion-induced subendocardial dark-rim artifacts. Step 1: the undersampled k-space from the perfusion image series is reconstructed to fully encoded k-space using the default scanner recon method, e.g., conventional GRAPPA with partial Fourier reconstruction. Step 2: The reconstructed fully sampled k-space is then used to generate two different recons by applying a temporal footprint (TF) reduction technique by removing consecutive phase-encode lines from one (left or right) side of the k-space and applying partial Fourier recon. Step 3: a spatiotemporal deep neural network (2D+time UNet-based architecture) is used to rapidly segment the myocardium into 12 sectors (standard AHA segmentation with endo-epi subsegments) and extract the average signal intensity (normalized by peak blood pool signal) for both left-sided and right-sided recons (shown as the top and bottom gray-scale bull’s eye maps, respectively). Next, the slope of the line connecting the left- vs right-sided S.I. for each segment is computed to generate a bull’s eye map of dark-rim TF variability (color-coded bull’s eye map). To automatically decide the optimal TF truncation (i.e., left vs right), the dark-rim TF variability is thresholded and combined with a “majority voting” approach (see the bull’s eye map with “L” and “R” votes) to automatically decide the optimal TF truncation (i.e., to choose the left vs right recon solution). The result is a DRA-suppressed images series (independently generated for each slice position) with minimal motion-induced DRA for each slice position.

Figure 2: Representative results for the proposed DRA elimination method. (A) These are two cases from the 10 stress FPP studies selected retrospectively based on “truly normal” invasive testing using a comprehensive protocol in the catheterization laboratory including: normal quantitative coronary angiography (no epicardial coronary stenosis > 20%), normal invasive coronary function testing (normal adenosine coronary flow reserve and normal coronary endothelial function), normal LV filling pressures and a negative LGE. The key feature of this validation approach is that it enables a notably more objective approach for DRA scoring since the stress scans are expected to be completely normal. As can be seen, in both cases, the scanner default recon has noticeable subendocardial DRA that can affect the assessment given the thin transmural extent (diastolic frame); however, the automatically generated optimal recon (green boxes) effectively eliminates the motion-induced DRA by selecting the optimal TF truncation approach (right-sided for Case 1 and left-sided for Case 2). In contrast, the non-optimal side results in a severe DRA (red arrows), which shows the impact of the TF choice (left vs right selection) on the presence and severity of DRA due to motion-induced (point spread function) effects. (B): Two examples from the feasibility testing of the inline implementation (FIRE framework) of the proposed technique. Both examples are clinically-indicated stress FPP studies on a 1.5T scanner (MAGNETOM Sola, Siemens Healthineers AG, Forchheim, Germany) wherein an additional image series is generated (besides the scanner default recon) with automatically generated DRA-optimal recon.

Figure 3: Summary of comparison between the scanner default recon and the proposed DRA-optimal recon in terms of (A) DRA severity, and (B) prevalence of severe DRA. By leveraging the availability of comprehensive invasive coronary angiography and coronary function testing in the selected subjects (n=10), the image artifact scores were assigned (DRA severity on 0-4 scale; 0: no artifact) by two expert readers in consensus. (A) DRA severity was significantly lower for the DRA-optimal approach vs default recon (1.5 ± 0.72 vs 2.2 ± 0.87, p <0.005) whereas the DRA severity for the non-optimal recon (i.e., selecting the opposite left/right side vs the optimal choice) is significantly higher than both (3.3 ± 0.74, p<0.001). (B) The percentage of slices with severe DRA (artifact score >2) was 47% for conventional method, 83% for non-optimal reconstruction and 7% for proposed DRA-optimal approach (p < 0.01 for all pairs. The proposed DRA-optimal reconstruction coupled with inline implementation provides an easily deployable DRA-reduction technique that can be applied to any standard- or high-resolution Cartesian-sampled FPP pulse sequence.