Segmentation model learning method, processing circuitry, computer program product, and medical information processing device
A segmentation model learning method according to an embodiment includes learning that, based on a loss function value, includes performing supervised learning of the voxels in medical image data according to the region to which the voxels belong. The learning of the medical image data includes: using first-type labeling information, which is meant for segmenting a predetermined structure into a plurality of categories, about the voxels of a predetermined structure and causing a segmentation model to perform direct supervised learning that represents learning for segmentation of the predetermined structure into a plurality of categories; using second-type labeling information, which is meant for segmenting a massive region covering the predetermined structure into a plurality of blocks, about the voxels of a massive region and causing the segmentation model to perform indirect supervised learning that represents learning for segmentation of the massive region into a plurality of categories; and optimizing the network parameters of the segmentation model.
1 . A segmentation model learning method based on weakly supervised learning, comprising:
obtaining, as learning data,
medical image data,
first-type labeling information meant for segmenting a predetermined structure into a plurality of categories, and
second-type labeling information meant for segmenting a massive region, which covers the predetermined structure, into a plurality of blocks; and
learning that, based on a loss function value, includes performing supervised learning of voxel in the medical image data according to a region to which the voxel belongs, wherein
the learning of the medical image data includes
using the first-type labeling information about voxel of the predetermined structure and causing a segmentation model to perform direct supervised learning that represents learning for segmentation of the predetermined structure in target medical image data for segmentation into a plurality of categories,
using the second-type labeling information about voxel of the massive region and causing the segmentation model to perform indirect supervised learning that represents learning for segmentation of the massive region in the target medical image data for segmentation into a plurality of categories, and
optimizing network parameter of the segmentation model.
2 . The segmentation model learning method according to claim 1 , wherein, the indirect supervised learning in the learning includes
obtaining, regarding each voxel of the massive region, maximum numerical value from among numerical values of concerned voxel in probability maps of a plurality of categories corresponding to a block to which concerned voxel belongs, and
performing supervised learning using the second-type labeling information regarding the maximum numerical value.
3 . The segmentation model learning method according to claim 2 , wherein, in the indirect supervised learning, when the maximum numerical value is equal to “1” and when maximum numerical value in a probability map of a single category from among a plurality of categories corresponding to a block to which concerned voxel belongs is equal to “1”, supervised learning is performed in such a way that numerical value of concerned voxel becomes equal to “0” in a probability map of other category other than the single category.
4 . The segmentation model learning method according to claim 1 , wherein the learning further includes learning labeling consistency of proximal voxels in such a way that the proximal voxels have identical labeling.
5 . The segmentation model learning method according to claim 1 , wherein the predetermined structure is a tubular region.
6 . The segmentation model learning method according to claim 5 , wherein the tubular region is a blood vessel or a trachea.
7 . The segmentation model learning method according to claim 5 , wherein the massive region is either one of a lung, a pulmonary lobe, liver, and hepatic lobe.
8 . A medical information processing device comprising:
an obtaining unit that obtains
target medial image data for segmentation which is collected from an examination target, and
an already-learnt segmentation model;
a processing unit that
segments the target medical image data for segmentation using the already-learnt segmentation model, and
obtains a segmentation result indicating segmentation of a predetermined structure and a massive region, which covers the predetermined structure, into a plurality of categories; and
an output unit that outputs the segmentation result, wherein
the already-learnt segmentation model is a segmentation model learnt according to the segmentation model learning method according to claim 1 .
9 . Processing circuitry that
obtains, as learning data,
medical image data,
first-type labeling information meant for segmenting a predetermined structure into a plurality of categories, and
second-type labeling information meant for segmenting a massive region, which covers the predetermined structure, into a plurality of blocks; and
performs learning that, based on a loss function value, includes
performing supervised learning of voxel in the medical image data according to a region to which the voxel belongs and by
using the first-type labeling information about voxel of the predetermined structure and causing a segmentation model to perform direct supervised learning that represents learning for segmentation of the predetermined structure in target medical image data for segmentation into a plurality of categories, and
using the second-type labeling information about voxel of the massive region and causing the segmentation model to perform indirect supervised learning that represents learning for segmentation of the massive region in the target medical image data for segmentation into a plurality of categories, and
optimizing network parameter of the segmentation model.
10 . A computer program product having a non-transitory computer-readable medium including programmed instructions, wherein the instructions, when executed by a computer, cause the computer to perform:
obtaining, as learning data,
medical image data,
first-type labeling information meant for segmenting a predetermined structure into a plurality of categories, and
second-type labeling information meant for segmenting a massive region, which covers the predetermined structure, into a plurality of blocks; and
learning that, based on a loss function value, includes performing supervised learning of voxel in the medical image data according to a region to which the voxel belongs, wherein
the supervised learning of the medical image data includes
using the first-type labeling information about voxel of the predetermined structure and causing a segmentation model to perform direct supervised learning that represents learning for segmentation of the predetermined structure in target medical image data for segmentation into a plurality of categories,
using the second-type labeling information about voxel of the massive region and causing the segmentation model to perform indirect supervised learning that represents learning for segmentation of the massive region in the target medical image data for segmentation into a plurality of categories, and
optimizing network parameter of the segmentation model.