Systems and methods for pseudo image data augmentation for training machine learning models
Systems and methods for augmenting a training data set with annotated pseudo images for training machine learning models. The pseudo images are generated from corresponding images of the training data set and provide a realistic model of the interaction of image generating signals with the patient, while also providing a realistic patient model. The pseudo images are of a target imaging modality, which is different than the imaging modality of the training data set, and are generated using algorithms that account for artifacts of the target imaging modality. The pseudo images may include therein the contours and/or features of the anatomical structures contained in corresponding medical images of the training data set. The trained models can be used to generate contours in medical images of a patient of the target imaging modality or to predict an anatomical condition that may be indicative of a disease.
1. A method for generating pseudo images of a target imaging modality, comprising:
selecting a pseudo-image generating process from a set of pseudo-image generating processes; and
applying the selected pseudo-image generating process on a first data set of medical images,
the medical images of the first data set including segmentation data and/or anatomical features related to an anomaly or disease,
the medical images of the first data set being of a first imaging modality,
the selecting of the pseudo-image generating process being by a user or an augmentation engine, and
the applying of the selected pseudo-image generating process includes applying one or more algorithms and process steps configured to generate a second data set of medical images using the medical images of the first data set and the segmentation data and/or anatomical features contained therein,
the medical images of the generated second data set including pseudo medical images of the target imaging modality that is different from the first imaging modality of the medical images of the first data set,
wherein the pseudo medical images are obtained by:
applying the selected pseudo-image generating process on the medical images of the first data set to generate pseudo image data of the target imaging modality; and
combining the medical images of the first data set with the generated pseudo image data of the target imaging modality.
2. The method of claim 1 , wherein the set of pseudo-image generating processes includes at least a forward projection process, a backward projection process, and a neural network process.
3. The method of claim 2 ,
wherein applying the forward projection process includes generating projection images from the medical images of the first data set,
wherein applying a backward projection process includes reconstructing the generated projection images into a pseudo volumetric image of the medical images, and
wherein applying a neural network process includes using a neural network trained to predict pseudo images of the target imaging modality based on images of the first imaging modality.
4. The method of claim 3 , wherein the generating of the pseudo images of the second data set includes applying a forward projection process combined with a backward projection process.
5. The method of claim 4 , comprising:
applying a forward projection algorithm on the medical images of the first data set to obtain projection images;
applying a back-projection algorithm on the generated projection images to accumulate the generated projection images onto a three-dimensional volume, the three-dimensional volume being a pseudo three-dimensional volumetric image representing a three-dimensional image of the target imaging modality; and
representing the pseudo three-dimensional volumetric image as a series of pseudo two-dimensional images or image slices.
6. The method of claim 5 , wherein the obtaining of the projection images includes:
obtaining a three-dimensional array of voxels from the medical images of the first data set, the array of voxels representing the medical images of the first data set viewed as a three-dimensional voxel volume;
applying an algorithm to simulate a hypothetical radiation beam from a virtual radiation source passing through the three-dimensional voxel volume at different radiation source positions and corresponding projection angles and captured on a plane of a virtual detector; and
generating line integrals across the radiation beams traversing the three-dimensional voxel volume to construct each of the projection images,
wherein the projection images represent projections of segmentation data and/or anatomical features of the three-dimensional voxel volume.
7. The method of claim 6 , wherein the forward projection algorithm takes into account radiation source and detector geometry used in the target imaging modality, geometry of the radiation beam in the target imaging modality, and one or more particle-matter interactions in the target imaging modality, in order to accurately simulate conditions present in the target imaging modality.
8. The method of claim 7 , wherein the one or more particle interactions include scatter and penumbra present in the target imaging modality, and wherein the forward projection algorithm includes parameters for scatter correction.
9. The method of claim 5 , wherein the accumulating of the generated projection images takes into consideration angles from which the projection images were acquired, and the generated three-dimensional pseudo volumetric image contains the segmentation data and/or anatomical features of the medical images of the first data set.
10. The method of claim 9 , wherein the back-projection algorithm accurately models behavior of radiation beams passing and scattering through the three-dimensional pseudo volumetric image to accurately simulate the conditions present in the target imaging modality.
11. The method of claim 10 , wherein the back-projection algorithm includes an iterative reconstruction algorithm including parameters to correct artifacts introduced by one or more of image lag, detector scatter, radiation beam hardening, motion.
12. The method of claim 3 , wherein the generating of the pseudo images of the second data set includes applying a forward projection process combined with a neural network process.
13. The method of claim 12 , comprising:
applying a forward projection algorithm on the medical images of the first data set to obtain projection images;
inputting the projection images into a neural network model to train the neural network model to generate an output, the output including segmentation data and/or anatomical features of the projection images; and
applying the trained neural network model on the medical images of the first data set to obtain segmentation and/or anatomical features data of the target imaging modality; and
combining the medical images of the first data set with the obtained segmentation and/or anatomical features data to obtain pseudo images of the target imaging modality.
14. The method of claim 3 , wherein the generating of the pseudo images of the second data set includes applying a neural network process.
15. The method of claim 14 , comprising:
selecting a neural network process trained to automatically generate a pseudo image of a target imaging modality from an image of a first imaging modality; and
applying the trained neural network process on the medical images of the first data set to generate the pseudo image data of the target imaging modality.
16. A system comprising:
one or more data storage devices storing a set of pseudo-image generating processes; and
one or more processors configured to:
receive a first data set of medical images including segmentation data and/or anatomical features related to an anomaly or disease, the medical images of the first data set being of a first imaging modality,
select a pseudo-image generating process to be applied on the medical images of the first data set from the set, and
apply the selected pseudo-image generating process on the medical images of the first data set to generate a second data set including pseudo-images of a target imaging modality,
wherein the applying the selected pseudo-image generating process includes applying one or more algorithms and process steps configured to generate the second data set of medical images using the medical images of the first data set and the segmentation data and/or anatomical features contained therein,
wherein the target imaging modality of the pseudo-images of the second data set is different from the first imaging modality of the medical images of the first data set, and
wherein the pseudo medical images are obtained by:
applying the selected pseudo-image generating process on the medical images of the first data set to generate pseudo image data of the target imaging modality; and
combining the medical images of the first data set with the generated pseudo image data of the target imaging modality.
17. The system of claim 16 , wherein the set of pseudo-image generating processes includes at least a forward projection process, a backward projection process, and a neural network process,
wherein applying the forward projection process includes generating projection images from the medical images of the first data set,
applying a backward projection process includes reconstructing the generated projection images into a pseudo volumetric image of the medical images,
applying a neural network process includes using a neural network trained to predict pseudo images of the target imaging modality based on images of the first imaging modality,
wherein the generating of the pseudo images of the second data set includes one of applying a forward projection process combined with a backward projection process, applying a forward projection process combined with a neural network process; and applying a trained neural network process.
18. The system of claim 17 , wherein the applying a forward projection process combined with a backward projection process includes:
applying a forward projection algorithm on the medical images of the first data set to obtain projection images;
applying a back-projection algorithm on the generated projection images to accumulate the generated projection images onto a three-dimensional volume, the three-dimensional volume being a pseudo three-dimensional volumetric image representing a three-dimensional image of the target imaging modality; and
representing the pseudo three-dimensional volumetric image as a series of pseudo two-dimensional images or image slices.
19. The system of claim 17 , wherein applying a forward projection process combined with a neural network process includes:
applying a forward projection algorithm on the medical images of the first data set to obtain projection images;
inputting the projection images into a neural network model to train the neural network model to generate an output, the output including segmentation data and/or anatomical features of the projection images; and
applying the trained neural network model on the medical images of the first data set to obtain segmentation and/or anatomical features data of the target imaging modality; and
combining the medical images of the first data set with the obtained segmentation and/or anatomical features data to obtain pseudo images of the target imaging modality.
20. The system of claim 17 , wherein applying a trained neural network process includes:
selecting a neural network process trained to automatically generate a pseudo image of a target imaging modality from an image of a first imaging modality; and
applying the trained neural network process on the medical images of the first data set to generate the pseudo image data of the target imaging modality.