Deep Learning for cardiac segmentation
Deep Learning for Cardiac images Segmentation: semi-automatic method to build Left Ventricle mesh
Project Outline
- 2D automatic segmentation for Short Axis MRI Endocardium
- 2D automatic segmentation for S-A MRI Myocardium
- 3D Mesh building and shape refining
- 4D segmentation
All the details and results can be found in the presentation:
Short-Axis MRI Dataset
Two datasets used:
- Sunnybrook dataset , 100 patients, 2 different time (ED and ES) -> 200 3D patients images -> transformed in 2D slice images. Contour: Inner endocardium of Left Ventricle (and also outer epicardium for ED)
- ACDC Dataset , (result in) 95 patient at ED and ES -> 190 3D patients images. Contour: Inner LV, Outer LV and RV
Below an example from Sunnybrook data. We can see all the patient's slices through the short axis and the relatively mask of the Inner endocardium of LV
ACDC Dataset provide also 4D images, with the addition of temporal dimension over a whole cardiac cycle
First Model: segmentation on Sunnybrook and ACDC separately
All the models are built using the python library "Segmentation Models"
- Segmentation of LV Endocardium on sunnybrook dataset using UNet with VGG16
- Segmentation of LV Endocardium, LV Epicardium and RV on ACDC dataset using UNet with VGG16
UNet model architecure
Ring Model: segmentation on Sunnybrook and ACDC together
Segmentation of LV Myocardium on a dataset with both Sunnybrook and ACDC data
- Random-sized crop
- Horizontal and Vertical flip
- Random rotation with an angle in [-90°,90°]
- Blur the image
- Random change of brightness
Example of generated augmented images and relatively augmented masks
Result:
3D Mesh building and Shape Refinement
Pipeline for shape refinement and building volume:
- Extract the endo and epi from ring images using OpenCV
- Extract contours of both endo and epi from the slices
- (Optionally) Shift correction for the contours
- Create the volume
- Add the apex if required
All the code for points 2 to 5 comes from https://github.com/cbutakoff/tools/tree/master/Python/mri_mesh_from_contours
Result examples from a test set patient
Comparison between true volume and predicted with our Cardiac Deep Learning
4D Segmentation
Finally, we can segment the volume of LV during a complete cardiac cycle to see the cardiac dynamic of the Ventricle. Below an example from ACDC Dataset: