Determination of a growth rate of an object in 3D data sets using deep learning

1. A method for automated determination of a growth rate of a first abnormality from one or more abnormalities in a body part of a patient, the method comprising:

a processor providing a current 3D data set associated with a first time instance to a first 3D deep neural network (DNN) system, the current 3D data set defining a first voxel representation of the body part of the patient, the first 3D DNN system being trained (i) to receive the current 3D data set, (ii) to determine whether the current 3D data set comprises the one or more abnormalities and (iii), if the current 3D data set comprises the one or more abnormalities, to output one or more first volumes of interest (VOIs) in the current 3D data set, each of the one or more first VOIs being associated with a respective abnormality from the one or more abnormalities;

the processor receiving the one or more first VOIs in the current 3D data set from the first 3D DNN system;

the processor providing a prior 3D data set associated with a second time instance to the first 3D DNN system, the prior 3D data set defining a second voxel representation of the body part of the patient;

the processor receiving one or more second VOIs in the prior 3D data set from the first 3D DNN system;

the processor using a registration algorithm to register the one or more first VOIs with the one or more second VOIs, the registration algorithm generating a mapping, the mapping determining, for a first VOI in the current 3D data set, a corresponding second VOI in the prior 3D data set, the first VOI and the corresponding second VOI being associated with a same abnormality from the one or more abnormalities;

the processor providing voxels of the first VOI a second 3D DNN system, the second 3D DNN system being trained (i) to receive voxels of the first VOI and (ii) to output a first 3D map defining probabilities for the voxels of the first VOI, a probability associated with a voxel defining a chance that the voxel is part of the one or more abnormalities;

the processor receiving the first 3D map from the second 3D DNN system;

the processor providing voxels of the corresponding second VOI to the second 3D DNN system;

the processor receiving a second 3D map from the second 3D DNN system;

the processor using (i) the first 3D map to identify first voxels in the current 3D data set representing the first abnormality and (ii) the second 3D map to identify second voxels in the prior data set representing the first abnormality; and;

the processor determining a first volume of the first abnormality based on the first voxels and a second volume of the first abnormality based on the second voxels and using the first volume and second volume to determine the growth rate of the first abnormality from the one or more abnormalities in the body part of the patient.

2. The method according to claim 1 , further comprising:

after receiving the one or more first VOIs, the processor using metadata associated with the current 3D data set to construct a request message and to send the request message to a database, the request message instructing the database to send the prior 3D data set to the processor.

3. The method according to claim 1 , wherein the first 3D DNN system includes:

at least a first 3D deep convolutional neural network (CNN), the first 3D deep CNN being trained to receive the current 3D data set and to output locations within the current 3D data set of one or more candidate VOIs, each candidate VOI defining a location in the current 3D data set at which the abnormality may be present; and

at least a second 3D deep CNN, the second 3D deep CNN being trained to receive a candidate VOI from the one or more candidate VOIs from the first 3D deep CNN and to determine a probability that voxels of the candidate VOI represent the abnormality.

4. The method according to claim 1 , wherein the registration algorithm includes a non-rigid transform to register voxels of the current 3D data set with voxels of the prior 3D data set.

5. The method according to claim 1 , wherein the registration algorithm comprises a third 3D DNN system trained to receive the first VOI of the current 3D data set and the second VOI of the prior 3D data set and to output a similarity score, the similarly score defining a measure regarding similarity between voxels of the first VOI and voxels of the second VOI.

6. The method according to claim 5 , wherein the registration algorithm generating the mapping includes:

determining a similarity matrix comprising probability scores associated with combinations of a first VOI selected from the one or more first VOIs in the current 3D data set and a second VOI selected from the one or more second VOIs in the prior 3D data set; and

using a linear optimization algorithm based on the similarity matrix to determine an optimal mapping between the one or more first VOIs of the current 3D data set and the one or more second VOIs of the prior 3D data set.

7. The method according to claim 5 , wherein the first 3D DNN system or the third 3D DNN system comprises a 3D residual convolutional neural network.

8. The method of claim 5 , wherein the third 3D DNN system is configured as a 3D deep Siamese neural network, the 3D deep Siamese neural network including a first 3D deep neural network part for receiving and processing the first VOI and a second 3D deep neural network part, wherein the first and second 3D deep neural network parts share same weights.

9. The method according to claim 1 , wherein:

a first threshold is applied to the probabilities in first 3D map to form a first 3D binary map identifying the first voxels in the current 3D data set; and

a second threshold is applied to the probabilities in the second 3D map to form a second 3D binary map to identify the second voxels in the prior 3D data set.

10. The method of claim 9 , wherein the first threshold is selected such that the sum of voxel volumes identified by the first 3D binary map represents the volume of the abnormality in the current 3D data set and the second threshold is selected such that the sum of voxel volumes identified by the second 3D binary map represents the volume of the abnormality in the prior 3D data set.

11. The method according to claim 1 , wherein the method further comprises a step of:

generating a digital report associated with the current 3D data set and the prior 3D data set, the digital report including a 3D graphical representation of the abnormality in the current 3D data set and a 3D graphical representation of the abnormality in the prior 3D data set and the growth rate of the abnormality.

12. The method according to claim 1 , wherein the first 3D DNN system and/or the second 3D DNN system comprises one or more blocks of convolutional layers, each block including a 3D CNN and a 2D CNN, wherein a reshaping operation reshapes slices of the 3D CNN into a plurality of 2D slices, wherein each 2D slice is processed by the 2D CNN.

13. The method according to claim 1 , wherein the storage and the retrieval of the current and prior 3D data sets are based on a DICOM standard.

14. Computer program product stored on a non-transitory computer-readable medium, the computer program product comprising software code portions configured for, when run in the memory of a computer, executing the method according to claim 1 .

15. The method of claim 1 , wherein the current 3D data set and the prior 3D data set are CT scans of the body part of the patent.

16. A method of training a plurality of 3D deep neural networks (DNNs), the method comprising the steps of:

receiving a training set for training a plurality of 3D DNNs, the training set including 3D data sets, wherein each of the 3D data sets either comprise zero or one or more abnormalities and one or more volume of interests (VOIs) for at least part of the 3D data sets, each VOI being associated with an abnormality;

receiving, for each VOI, a pixel representation of the VOI, location information indicating at which location the VOI is located in a 3D data set, and a probabilistic 3D map defining probabilities for voxels of the VOI, a probability associated with a voxel defining a chance that the voxel is part of an abnormality;

training a first 3D DNN using voxel representations of the 3D data sets as input and the location information as a target;

training a second 3D DNN using voxel representations of the VOIs as input and the probabilistic 3D map associated with the voxel representations of the VOIs as the target; and

training a third 3D DNN using voxel representations of the VOIs and non-linear image transformations of the voxel representations of the VOIs as input and similarity scores as the target, a similarity score defining a similarity between a voxel representation of a VOI and a non-linear image transformation of the voxel representation.

17. The method of claim 16 , wherein the current 3D data set and the prior 3D data set are CT scans, of the body part of the patent.

18. A computer system for automated determination of a growth rate of a first abnormality from one or more abnormalities in a body part of a patient, the computer system comprising:

a non-transitory computer-readable storage medium having computer-readable program code embodied therewith, the program code including at least one trained 3D deep neural network, and

at least one processor coupled to the computer-readable storage medium, wherein, responsive to executing the computer-readable program code, the at least one processor is configured to perform executable operations comprising:

providing a current 3D data set associated with a first time instance to a first 3D deep neural network (DNN) system, the current 3D data set defining a first voxel representation of the body part of the patient, the first 3D DNN system being trained (i) to receive the current 3D data set, (ii) to determine whether the current 3D data set comprises the one or more abnormalities and (iii), if the current 3D data set comprises the one or more abnormalities, to output one or more first volumes of interest (VOIs) in the current 3D data set, the one or more VOIs being associated with the a respective abnormality from the one or more abnormalities;

receiving the one or more first VOIs in the current 3D data set from the first 3D DNN system;

providing a prior 3D data set associated with a second time instance to the first 3D DNN system, the prior 3D data set defining a second voxel representation of the body part of the patient;

receiving one or more second VOIs in the prior 3D data set from the first 3D DNN system;

using a registration algorithm to register the one or more first VOIs with the one or more second VOIs, the registration algorithm generating a mapping, the mapping determining, for a first VOI in the current 3D data set, a corresponding second VOI in the prior 3D data set, the first VOI and the corresponding second VOI being associated with a same abnormality from the one or more abnormalities;

providing voxels of the first VOI to a second 3D DNN system, the second 3D DNN system being trained to receive voxels of the first VOI and to output a first 3D map defining probabilities for voxels of the first VOI, a probability associated with a voxel defining a chance that the voxel is part of the one or more abnormalities;

receiving the first 3D map from the second 3D DNN system;

providing voxels of the corresponding second VOI to the second 3D DNN system;

receiving a second 3D map from the second 3D DNN system;

using the first 3D map to identify first voxels in the current 3D data set representing the first abnormality;

using the second 3D map to identify second voxels in the prior data set representing the first abnormality; and

determining a first volume of the first abnormality based on the first voxels and a second volume of the first abnormality based on the second voxels and using the first volume and second volume to determine the growth rate of the first abnormality from the one or more abnormalities in the body part of the patient.

19. The computer system according to claim 18 , wherein the first 3D DNN system includes:

at least a first 3D deep convolutional neural network (CNN), the first 3D deep CNN being trained to receive the current 3D data set and to output locations within the current 3D data set of one or more one candidate VOIs, each candidate VOI defining a location in the current 3D data set at which an abnormality may be present; and

at least a second 3D deep CNN, the second 3D deep CNN being trained to receive a candidate VOI from the one or more candidate VOIs from the first 3D deep CNN and to determine a probability that voxels of the candidate VOI represent the abnormality.

20. The computer system of claim 19 , wherein the third 3D DNN system is configured as a 3D deep Siamese neural network, the 3D deep Siamese neural network including a first 3D deep neural network part for receiving and processing the first VOI and a second 3D deep neural network part, wherein the first and second 3D deep neural network parts share same weights.

21. The computer system according to claim 18 , wherein the registration algorithm comprises a third 3D DNN system trained to receive the first VOI of the current 3D data set and the second VOI of the prior 3D data set at its input and to determine a similarity score at its output, the similarly score defining a measure regarding similarity between voxels of the first VOI and voxels of the second VOI.