Deep learning-based eddy current correction
Systems and methods for providing improved eddy current correction (ECC) in medical imaging environments. One or more of the embodiments disclosed herein provide a deep learning-based convolutional neural network (CNN) model trained to automatically generate an ECC mask which may be composited with two-dimensional (2D) scan slices or four-dimensional (4D) scan slices and made viewable through, for example, a web application, and made manipulable through a user interface thereof.
1. An image processing system, comprising:
at least one nontransitory processor-readable storage medium storing at least one of processor-executable instructions or data, the data including magnetic resonance imaging (MRI) image data; and
at least one processor communicatively coupled to the at least one nontransitory processor-readable storage medium, in operation the at least one processor:
accesses the MRI image data;
calculates, based on a trained convolutional neural network (CNN) model, a segmentation mask applicable to the MRI image data, wherein the segmentation mask differentiates static tissue and non-static tissue, wherein the trained CNN model was trained using a loss function to compare generated eddy current distortion segmentation masks for image slices with corresponding pre-generated eddy current correction mask annotations for the image slices; and
applies, using a polynomial fitting function, the segmentation mask to the MRI image data to produce a corrected image which accounts for eddy current distortion in the MRI image data.
2. The image processing system of claim 1 , wherein the trained CNN model was trained using a differentiable loss function comprising weighted and unweighted cross entropy, weighted and unweighted L 2 loss, log dice index, or log jaccard index.
3. The image processing system of claim 1 , wherein the at least one processor applies one or more post-processing techniques to the calculated segmentation mask comprising conditional random fields, Gaussian blurring, dilation, erosion, hole filling, histogram equalization, or Gaussian convolving.
4. The image processing system of claim 1 , wherein the at least one processor provides an adjustable probability threshold cutoff of the segmentation mask to select between one or more interpretations of the MRI image data, each interpretation associated with a tensor comprising per-pixel probability of representing static tissue or non-static tissue.
5. The image processing system of claim 1 , wherein the segmentation mask is applicable to cardiac image data at any point throughout a cardiac cycle.
6. The image processing system of claim 1 , wherein the segmentation mask is applicable to MRI image data comprising axial data, sagittal data, or coronal data.
7. The image processing system of claim 1 , wherein the accessed MRI image data comprises at least one of 2D flow data or 4D flow data.
8. The image processing system of claim 1 , wherein the trained CNN model is configured to calculate segmentations masks for 2D, 3D, or 4D MRI image data.
9. The image processing system of claim 1 , wherein the trained CNN model is trained based on one or more training data inputs comprising 2D MRI image data, 3D MRI image data, or 4D MRI image data.
10. The image processing system of claim 9 , wherein the dimensionality of the MRI image data to which the segmentation mask is applicable is greater than or equal to the dimensionality of the one or more training data inputs.
11. The image processing system of claim 1 , wherein the trained CNN model is configured to generate segmentation masks applicable to 2D MRI image data, 3D MRI image data, or 4D MRI image data.
12. The image processing system of claim 1 , wherein the trained CNN model was trained using a differentiable loss function comprising cross entropy or L 2 loss.
13. A method of processing image data, the method comprising:
accessing, by at least one processor, magnetic resonance imaging (MRI) image data;
calculating, by the at least one processor and based on a trained convolutional neural network (CNN) model, a segmentation mask applicable to the MRI image data, wherein the segmentation mask differentiates static tissue and non-static tissue, wherein the trained CNN model was trained using a loss function to compare generated eddy current distortion segmentation masks for image slices with corresponding pre-generated eddy current correction mask annotations for the image slices; and
applying, by the at least one processor and using a piecewise linear fitting function, the segmentation mask to the MRI image data to produce a corrected image which accounts for eddy current distortion in the MRI image data.
14. The method of claim 13 , wherein the trained CNN model was trained using a differentiable loss function comprising weighted and unweighted cross entropy, weighted and unweighted L 2 loss, log dice index, or log jaccard index.
15. The method of claim 13 , further comprising applying one or more post-processing techniques to the calculated segmentation mask comprising conditional random fields, Gaussian blurring, dilation, erosion, hole filling, histogram equalization, or Gaussian convolving.
16. The method of claim 13 , further comprising providing an adjustable probability threshold cutoff of the segmentation mask to select between one or more interpretations of the MRI image data, each interpretation associated with a tensor comprising per-pixel probability of representing static tissue or non-static tissue.
17. The method of claim 13 , wherein the segmentation mask is applicable to cardiac image data at any point throughout a cardiac cycle.
18. The method of claim 13 , wherein the segmentation mask is applicable to MRI image data comprising axial data, sagittal data, or coronal data.
19. The method of claim 13 , wherein accessing the MRI image data comprises accessing at least one of 2D flow data or 4D flow data.
20. The method of claim 13 , wherein the trained CNN model is configured to calculate segmentations masks for 2D, 3D, or 4D MRI image data.
21. The method of claim 13 , wherein the trained CNN model was trained based on one or more training data inputs comprising 2D MRI image data, 3D MRI image data, or 4D MRI image data.
22. The method of claim 21 , wherein the dimensionality of the MRI image data to which the segmentation mask is applicable is greater than or equal to the dimensionality of the one or more training data inputs.
23. The method of claim 13 , wherein the trained CNN model is configured to generate segmentation masks applicable to 2D MRI image data, 3D MRI image data, or 4D MRI image data.
24. The method of claim 13 , wherein the trained CNN model was trained using a differentiable loss function comprising log dice index or log jaccard index.