1351770 - Extending DENSEanalysis for automated and faster DENSE strain processing

Myocardial strain analysis plays a key role in assessing cardiovascular conditions. Cine Displacement ENcoding with Stimulated Echoes (DENSE) encodes displacement in the image phase, enabling accurate and high-resolution strain extraction [1, 2].

Strain processing with DENSE is done using standardized tools like DENSEanalysis [3], which require a lot of user intervention. For DENSE to be more widely available in clinical settings, it is necessary to reduce processing times and mitigate user-related variability.  

The objective of the present research is to accelerate short and long axis DENSEanalysis processing. We extend the existing software to reduce user interaction and enable the analysis of full datasets with minimal processing time.

Methods: Segmentation of the left ventricular myocardium is the DENSEanalysis step that requires the most user interaction. To accelerate it, we trained a nnU-Net-based Deep Learning (DL) model, and used it in inference mode to process remaining short and long axis datasets. Data loading and result exporting were automated and any graphical user interface (GUI) was removed. Segmentation masks were converted into epi and endocardial contours and uploaded as region of interest (ROI) entries on DENSEanalysis workspaces. For phase unwrapping, smoothing parameters were defined beforehand, and seed points in non-wrapped regions were pre-set by positioning points at standardized locations on frame 4 of the cine sequence. The only user interaction left was to adjust for pre-set seed points if needed and adjust the right ventricular (RV) insertion point on the AHA segmental model for strain analysis of short-axis cases. Remaining user interaction and GUI can be seen in Fig 2. The final pipeline was scripted for the rapid processing of large datasets.

Results: After having trained the segmentation model (21 hours on an NVIDIA GeForce RTX 3090 GPU with 24 GB RAM), inference estimation and segmentation loading on DENSEanalysis takes around a minute. Phase unwrapping, strain extraction, and results exporting can be performed in approximately 30 seconds. The modified pipeline substantially reduces the analysis time, which, manually, can take up to 40 minutes per case from data loading to strain results. Out of 100 cases (short and long-axis combined), seed points for phase unwrapping had to be slightly modified for 5 cases, due to some of the pre-set seed points being placed on wrapped pixels. Fig 3 shows that the automatic circumferential and longitudinal strain was similar to the manual processing pipeline, while radial strain agreement remains within the limits of intra-user variability [4].

Conclusion: The automation of DENSEanalysis provides an efficient pipeline for DENSE processing. It will enable the analysis of large datasets, and facilitate the development and integration of Deep Learning models in the pipeline [5]. In future work, we will extend this pipeline to include additional Deep Learning models, for phase unwrapping and strain extraction.