Material segmentation in image volumes
The present approach relates, in some aspects, to a multi-level and a multi-channel frame work for segmentation using model-based or “shallow” classification (i.e. learning processes such as linear regression, clustering, support vector machines, and so forth) followed by deep learning. This framework starts with a very low resolution version of the multi-channel data and constructs an shallow classifier with simple features to generate a coarser level tissue mask that in turn is used to crop patches from the high-resolution volume. The cropped volume is then processed using the trained convolution network to perform a deep learning based segmentation within the slices.
1. A method, comprising:
obtaining high resolution multi-channel measurement data from a medical imaging scanner;
downsampling high resolution multi-channel measurement data to a low resolution volume;
using a model based or shadow classification to generate a low resolution tissue mask from the low resolution volume;
cropping patches from high resolution multi-channel measurement data corresponding to the low resolution tissue mask;
generating 2D slices from the cropped patches;
processing 2D slices using a trained neural network, wherein the trained neural network outputs a set of two-dimensional or higher dimensional single-channel tissue maps;
generating a tissue mask using the set of single-channel tissue maps; and
displaying the tissue mask or a tissue-free volume derived using the tissue mask.
2. The method of claim 1 , wherein the tissue map comprises a bone map and the tissue mask comprises a bone mask.
3. The method of claim 1 , wherein the high resolution multi-channel measurement data comprises dual-energy or higher energy simultaneously acquired or co-registered measurement data.
4. The method of claim 1 , wherein the single-channel tissue maps comprise tissue probability maps and generating the tissue mask comprises:
threshold processing the tissue probability maps to generate the tissue mask slices and/or volume.
5. The method of claim 1 , wherein the trained neural network comprises a plurality of downsampling layers and a corresponding plurality of upsampling layers.
6. The method of claim 1 , wherein the high resolution multi-channel measurement data is co-registered.
7. An image processing system comprising:
a processor configured to execute one or more stored processor-executable routines; and
a memory storing the one or more executable-routines, wherein the one or more executable routines, when executed by the processor, cause acts to be performed comprising:
obtaining high resolution multi-channel measurement data from one or more medical imaging scanners;
downsampling high resolution multi-channel measurement data to a low resolution volume;
using a model based or shadow classification to generate a low resolution tissue mask from the low resolution volume;
cropping patches from high resolution multi-channel measurement data corresponding to the low resolution tissue mask;
generating 2D slices from the cropped patches;
processing 2D slices using a trained neural network, wherein the trained neural network outputs a set of two or more-dimensional single-channel tissue maps;
generating a tissue mask using the set of single-channel tissue maps; and
displaying the tissue mask or a target tissue-free volume derived using the tissue mask.
8. The image processing system of claim 7 , wherein the high resolution multi-channel measurement data comprises dual-energy, multi-energy or multi-modality measurement data.
9. The image processing system of claim 7 , wherein the tissue map comprises a bone map and the tissue mask comprises a bone mask.
10. The image processing system of claim 7 , wherein the single-channel tissue maps comprise tissue probability maps and generating the tissue mask comprises:
threshold processing the tissue probability maps to generate a set of tissue mask images.
11. The image processing system of claim 7 , wherein the trained neural network comprises a plurality of downsampling layers and a corresponding plurality of upsampling layers.
12. One or more non-transitory computer-readable media encoding processor-executable routines, wherein the routines, when executed by a processor, cause acts to be performed comprising:
obtaining high resolution multi-channel measurement data from one or more medical imaging scanners;
downsampling high resolution multi-channel measurement data to a low resolution volume;
using a model based or shadow classification to generate a low resolution tissue mask from the low resolution volume;
cropping patches from high resolution multi-channel measurement data corresponding to the low resolution tissue mask;
generating 2D slices from the cropped patches;
processing 2D slices using a trained neural network, wherein the trained neural network outputs a set of two- or higher dimensional single-channel tissue maps;
generating a tissue mask using the set of single-channel tissue maps; and
displaying the tissue mask or a tissue-free volume derived using the tissue mask.
13. The one or more non-transitory computer-readable media of claim 12 , wherein the high resolution multi-channel measurement data comprises dual- or multi-energy or multi-modal measurement data.
14. The one or more non-transitory computer-readable media of claim 12 , wherein the two- or higher dimensional single-channel tissue maps comprise tissue probability maps and generating the tissue mask comprises:
processing of the tissue probability maps to generate a set of tissue mask images.