/DeepStrain

A Deep Learning Workflow for the Automated Characterization of Cardiac Mechanics

Primary LanguageJupyter NotebookMIT LicenseMIT

[paper] [installation] [examples]





DeepStrain

DeepStrain: A Deep Learning Workflow for the Automated Characterization of Cardiac Mechanics.

✨ New Updates.

  • 🚀 Trained weights moved to Harvard Dataverse: https://doi.org/10.7910/DVN/XB6PEZ.
  • 🚀 Weights script updated for easy download: bash ./pretrained_models/download_model.sh
  • 🚀 Cardiac Motion Analysis Challenge (CMAC) dataset uploaded to Dataverse!!!
  • 🚀 CMAC script available for easy download: bash ./datasets/download_cmac.sh
  • 🚀 Code to read CMAC landmarks uploaded (./cmac) see demo.

Tensorflow implementation for cardiac segmentation, motion estimation, and strain analysis from balanced steady-state free-precession (bSSFP) cine MRI images.

Note: The current software works well with Tensorflow 2.3.1+.

UPDATE 2024: The trained moel is now available through the Harvard Dataverse: https://doi.org/10.7910/DVN/XB6PEZ

Evaluation

DeepStrain Tracking in Patients Across MRI Vendors

Short-axis bSSFP cine MRI images are shown at the mid-ventricle. To visualize tracking, myocardial contours of the endocardial (red) and epicardial (green) left ventricular wall defined at end diastole were deformed to end systole using displacement vectors based on DeepStrain. Vendor 1 = 1.5T (Achieva; Philips Medical Systems, Best, the Netherlands). Vendor 2 = 3T (MAGNETOM Vida; Siemens Healthcare, Erlangen, Germany).



Evaluation Relative to Feature Tracking in Patients


Measurements of global radial and circumferential strain derived from cvi42 feature tracking were compared to those derived with DeepStrain. Each dot represents a single subject (n = 580). Solid line shows linear regression analyses based on images acquired in Philips (blue) and Siemens (red) MRI scanners. On the left, the slopes (y = 1.06×, y = 1.04×) represent a -6% and -4% disagreement in measurements of global radial strain. The slopes (y = 0.99×, y = 1.00×) on the right represent a 1% and 0% disagreement in measurements of global circumferential strain.


Agreement Between cvi42 Feature Tracking and DeepStrain

Linear Regression Slope GRS GCS
Vendor 1 (n = 466) 1.06 (1.03, 1.08) 0.98 (0.98, 1.01)
Vendor 2 (n = 114) 1.04 (0.99, 1.09) 1.0 (0.96, 1.03)
Total (n = 580) 1.05 (1.03, 1.08) 0.99 (0.98, 1.01)
Pearson Correlation Coefficient (r) GRS GCS
Vendor 1 (n = 466) 0.85 0.91
Vendor 2 (n = 114) 0.83 0.88
Total (n = 580) 0.85 0.91

GRC = global radial strain; GCS = global circumferential strain. Vendor 1 = 1.5T (Achieva; Philips Medical Systems, Best, the Netherlands); Vendor 2 = 3T (MAGNETOM Vida; Siemens Healthcare, Erlangen, Germany). Data are the slope (95% confidence interval of the slope) and Pearson correlation coefficient (r).

Application


The study cohort consisted of 119 participants (35 ± 5 years, 50% male) including the control group with 30 subjects; RFG1 with 39 overweight subjects; RFG2 with 30 hypertensive subjects, including 13 (43%) with additional overweight; RFG3 with 20 T2DM subjects, including 11 (55%) with additional overweight, 1 (5%) with additional hypertension and 8 (40%) with both. Measures of (a) circumferential and (b) radial early-diastolic strain rate (SR) in controls and risk factor groups. Strain results are visualized as mean with 95% confidence interval. Post-hoc test by Bonferroni: * p < 0.05; ** p < 0.01; *** p < 0.001. RFG = risk factor group.



Publications

If you find DeepStrain or some part of the code useful, please cite as appropiate: