CNNT DB-LGE: free-breathing dark blood late enhancement imaging using the convolutional neural network transformer speeds acquisition by 50%

Abstract

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – National budget only. Main funding source(s): Supported in part by the Division of Intramural Research of the National Heart, Lung, and Blood Institute, National Institutes of Health (grants Z1A-HL006214-05 and Z1A-HL006242-02). Background Dark blood late gadolinium enhancement (DB-LGE) imaging shows superior delineation of myocardial infarction (MI), especially at the sub-endocardial boundary. Our previous study [1] developed a free-breathing DB-LGE with the single shot SSFP readout, phase sensitive inversion recovery (PSIR) reconstruction, and respiratory motion corrected averaging. To compensate the potential signal-to-noise ratio loss, our previous DB-LGE doubled the measurements, thereby increasing the acquisition time. Purpose In this study, we developed a deep learning image enhancement model using a novel neural network architecture called the convolutional neural network transformer (CNNT) to improve the image quality of DB-LGE and to reduce the acquisition time by decreasing the number of measurements. Methods A novel image enhancement model was developed using a novel network architecture called the Convolutional Neural Network Transformer (CNNT) proposed by us. This architecture is suitable for the 2D+Time CMR acquisition, by exploiting the temporal correlation between images over multiple averages. The evaluation was first retrospectively conducted on a cohort of 12 patients acquired with the original protocol [1] using the full 16 measurements. For every subject, a complete short-axis stack (typically 12 slices) was acquired to cover the entire left ventricular. The imaging data was reconstructed in three ways. Original: using all acquired 16 measurements. This is our base-line protocol. Original 50%: using only the first 8 measurements. CNNT 50%: using only the first 8 averages, but performing the CNNT deep learning image enhancement before MOCO PSIR reconstruction. Two experienced imaging researchers (PK and MF, >10 years of experience for both) 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). The CNNT DB-LGE was deployed to the MR scanner using the Gadgetron InlineAI [2]. Results Figure 1 gives examples of DB-LGE with three reconstruction methods. The CNNT image has higher SNR and well delineated MI. The Original images with the longest acquisition have good quality and the Original-50% acquired with 8 measurements are good quality but have reduced SNR. The mean scores for overall image quality, diagnostic confidence and 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 (P>0.15 for all). Figure 2 shows an acute MI patient prospectively acquired with the 50% scan time reduction, with and without the CNNT enhancement. The resulting PSIR images well delineate the MVO due to the acute MI, with improved SNR. Conclusion A novel CNNT model was proposed and evaluated to speed up the free-breathing MOCO DB LGE by 50% without sacrificing image quality.