1338633 - Dark blood PSIR LGE with AI denoising speeds acquisition by 50%

Dark blood late gadolinium enhancement (DB-LGE) imaging has excellent delineation of MI at the sub-endocardial boundary due to blood suppression [1]. DB LGE is acquired free-breathing with single shot readout, phase sensitive inversion recovery (PSIR) reconstruction, and respiratory motion corrected averaging. To compensate the signal-to-noise ratio loss due to the inversion recovery T2 preparation used to suppress the blood signal, our previous DB-LGE doubled the number of measurements, thereby increasing the acquisition time.

Methods:

The CNNT architecture utilizes the 2D+Time acquisition exploiting the temporal correlation between images over multiple measurements. Its design promotes the separation of the temporally correlated signal from uncorrelated random noise. The AI model was applied to both the inversion recovery and proton density weighted complex images. In this way, the downstream PSIR reconstruction can be performed and the total number of measurements can be reduced while achieving the desired SNR. For all approaches, in-plane respiratory MOCO is performed and through plane motion is mitigated by discarding 50% of the images based on a measure of dissimilarity. PSIR reconstruction is performed on the averaged images to compute the DB-PSIR LGE image.

A retrospective evaluation was performed in 12 patients using the original protocol [1] with 16 measurements. For each subject, a short-axis stack was acquired to cover the entire left ventrical and was reconstructed in three ways. Original: using all 16 measurements. Original 50%: using only the first 8 measurements. CNNT 50%: using the first 8 measurements with CNNT AI denoising before MOCO PSIR reconstruction. Two experienced imaging researchers (each with >10 years of experience) scored all DB-LGE images for the overall quality, diagnostic confidence and delineation of MI/boundaries (5 = excellent, 4 = good, 3 = fair, 2 = poor, and 1 = non-diagnostic). Images were randomized and the reviewer was blind to the processing method.

The CNNT DB-LGE was deployed to the MR scanner using Gadgetron InlineAI [2]. For a SAX stack, it saves 192 HBs (~3mins for 60 bpm) which is substantial time reduction. This study was approved by the local Ethics Committee, with written consent.

Results:

The mean scores for a) overall image quality, b) diagnostic confidence and c) MI delineation of two reviewers were 4.88±0.23, 4.88±0.23, 4.83±0.25 for CNNT and 4.96±0.14, 4.96±0.14, 4.67±0.39 for the original approach. No significant differences were found between the original and the CNNT 50% (P >0.15 for all).

Fig 1 gives examples of DB-LGE with three reconstruction methods. Fig 2 gives an example for an acute MI patient prospectively acquired with the proposed 50% scan time reduction, with and without the CNNT enhancement.

Conclusion:

A novel deep learning model named CNNT was proposed and evaluated to speed up the image acquisition time of free-breathing MOCO DB LGE by 50% without sacrificing image quality.