Systems and methods for segmentation of anatomical structures for image-guided surgery
A method for image segmentation comprises receiving volumetric image data for an anatomical region and generating a first volumetric patch from the volumetric image data. The method also comprises generating a second volumetric patch from the first volumetric patch by weighting a plurality of volumetric units in the first volumetric patch and receiving the second volumetric patch as an input to a convolutional neural network. The method also comprises conducting a down-sampling filter process and conducting an up-sampling filter process within the convolutional neural network.
1. A method for image segmentation, the method comprising:
receiving volumetric image data for an anatomical region;
generating a first volumetric patch from the volumetric image data;
generating a second volumetric patch from the first volumetric patch by weighting a plurality of volumetric units in the first volumetric patch using a multi-scale entropy map providing an indication of structural detail between scale levels;
receiving the second volumetric patch as an input to a convolutional neural network;
within the convolutional neural network, conducting a down-sampling filter process; and
within the convolutional neural network, conducting an up-sampling filter process.
2. The method of claim 1 wherein weighting at least one of the plurality of volumetric units includes applying a weight based on edge features identified in the multi-scale entropy map.
3. The method of claim 1 wherein weighting at least one of the plurality of volumetric units includes applying a weight based on structure complexity identified in the multi-scale entropy map.
4. The method of claim 1 wherein weighting at least one of the plurality of volumetric units includes applying a weight based on a distance from a volumetric unit, in the first volumetric patch, with a foreground structure classification.
5. The method of claim 1 wherein weighting the plurality of volumetric units includes applying a weight based on a foreground structure classification.
6. The method of claim 1 wherein weighting the plurality of volumetric units includes applying a weight based on a patch significance for the first volumetric patch.
7. The method of claim 1 wherein weighting the plurality of volumetric units includes
identifying a set of the plurality of volumetric units that have a background structure classification and
applying a weight to a randomly selected volumetric unit of the set of the plurality of volumetric units.
8. The method of claim 1 wherein conducting the down-sampling filter process includes applying a first filter at a first resolution scale to generate a first output and applying a second filter to the first output at a second resolution scale.
9. The method of claim 8 wherein conducting the up-sampling filter process includes applying a third filter to the first output at the first resolution scale and to an output of the second filter at the second resolution scale.
10. A non-transitory machine-readable medium comprising a plurality of machine-readable instructions which, when executed by one or more processors, are adapted to cause the one or more processors to perform a method comprising:
receiving volumetric image data for an anatomical region;
generating a first volumetric patch from the volumetric image data;
generating a second volumetric patch from the first volumetric patch by weighting a plurality of volumetric units in the first volumetric patch using a multi-scale entropy map providing an indication of structural detail between scale levels;
receiving the second volumetric patch as an input to a convolutional neural network;
within the convolutional neural network, conducting a down-sampling filter process; and
within the convolutional neural network, conducting an up-sampling filter process.
11. The non-transitory machine-readable medium of claim 10 wherein weighting at least one of the plurality of volumetric units includes applying a weight based on edge features identified in the multi-scale entropy map.
12. The non-transitory machine-readable medium of claim 10 wherein weighting at least one of the plurality of volumetric units includes applying a weight based on structure complexity identified in the multi-scale entropy map.
13. The non-transitory machine-readable medium of claim 10 wherein weighting at least one of the plurality of volumetric units includes applying a weight based on a distance from a volumetric unit, in the first volumetric patch, with a foreground structure classification.
14. The non-transitory machine-readable medium of claim 10 wherein weighting the plurality of volumetric units includes applying a weight based on a foreground structure classification.
15. The non-transitory machine-readable medium of claim 10 wherein weighting the plurality of volumetric units includes applying a weight based on a patch significance for the first volumetric patch.
16. The non-transitory machine-readable medium of claim 10 wherein weighting the plurality of volumetric units includes
identifying a set of the plurality of volumetric units that have a background structure classification and
applying a weight to a randomly selected volumetric unit of the set of the plurality of volumetric units.
17. The non-transitory machine-readable medium of claim 10 wherein conducting the down-sampling filter process includes applying a first filter at a first resolution scale to generate a first output and applying a second filter to the first output at a second resolution scale.
18. The non-transitory machine-readable medium of claim 17 wherein conducting the up-sampling filter process includes applying a third filter to the first output at the first resolution scale and to an output of the second filter at the second resolution scale.