Systems and methods for background aware reconstruction using deep learning
Method ( 1000 ) and system ( 100 ) for image processing for a medical device is provided. The method ( 1000 ) includes acquiring ( 1010 ) a plurality of images of a subject using an image acquisition system ( 110 ) of the medical device. The method ( 1000 ) further includes identifying and differentiating ( 1020 ) a plurality of background pixels and a plurality of foreground pixels in the images using the deep learning module ( 125 ). The method ( 1000 ) further includes suppressing ( 1030 ) the identified background pixels using a mask and processing the foreground pixels for subsequent reconstruction and/or visualization tasks.
1. A method for image processing for a magnetic resonance imaging (MRI) system, the method comprising:
receiving, via a deep learning module, a plurality of magnetic resonance images of a subject acquired using an image acquisition system of the MRI system;
differentiating a plurality of background pixels from a plurality of foreground pixels in the plurality of magnetic resonance images using the deep learning module, wherein the foreground pixels correspond to an anatomical region of the subject and the background pixels correspond to a non-anatomical and clinically irrelevant region located outside the anatomical region;
suppressing the background pixels using a mask by setting the values of only the background pixels to zero;
processing the foreground pixels to reconstruct higher image quality magnetic resonance images than the plurality of magnetic resonance images for visualization, wherein the plurality of magnetic resonance images and the higher image quality magnetic resonance images comprise localizer images;
storing the mask; and
utilizing the mask on magnetic resonance imaging data acquired in subsequent scans to generate diagnostic quality magnetic resonance images.
2. The method as claimed in claim 1 , further comprising generating the mask using the deep learning module to suppress the background pixels.
3. The method as claimed in claim 1 , further comprising generating one mask for each axial, sagittal and coronal orientations.
4. The method as claimed in claim 1 , wherein the suppressed background pixels are recovered later for image reconstruction and visualization.
5. The method as claimed in claim 1 , wherein processing the foreground pixels comprises performing morphological operations on the mask to recover any potential loss of anatomical regions.
6. The method as claimed in claim 1 , wherein the deep learning module comprises an inferencing module adapted to generate the mask.
7. The method as claimed in claim 1 , wherein a localizer scanner is adapted to generate localizer images that are processed by the deep learning module.
8. A system for acquiring and processing an image in a medical device, the system comprising:
an image acquisition system of the medical device configured to acquire magnetic resonance images of a subject;
an image reconstruction module connected to the image acquisition system and configured to receive magnetic resonance images from the image acquisition system;
a deep learning module connected to the image reconstruction system and adapted to generate a mask; and
a visualization module operatively coupled to the image reconstruction model and configured to further process the magnetic resonance images to generate display quality magnetic resonance images;
wherein the deep learning module is configured to identify and differentiate the background pixels from the foreground pixels in the magnetic resonance images, and the image reconstruction module is adapted to use this information during early stages of reconstruction to suppress background pixels using a mask that sets values of only the background pixels to zero and process only foreground pixels during subsequent steps of an image reconstruction pipeline for storage and post processing, wherein the foreground pixels correspond to an anatomical region of the subject and the background pixels correspond to a non-anatomical and clinically irrelevant region located outside the anatomical region, and wherein a fully reconstructed or DICOM image of a localizer is used to generate the mask and the mask is used for suppressing the background pixels in the localizer as well as diagnostic quality images reconstructed from magnetic resonance imaging data acquired in subsequent scans.
9. The system as claimed in claim 8 , wherein the deep learning module comprises an inferencing module and generates the mask.
10. The system as claimed in claim 8 , wherein the visualization module comprises a display screen configured to show background suppressed visuals of the subject.
11. The system as claimed in claim 8 , wherein the deep learning module is a Convolutional Neural Network employing U-Net.
12. The system as claimed in claim 8 , wherein the deep learning module comprises an inferencing module and a binary mask adapted to suppress the background pixels.
13. The system as claimed in claim 8 , wherein the mask is stored.