3D reconstruction from a limited number of 2D projections
AssigneesTherapanacea, Institut De Cancerologie Gustave Roussy, Centre Leon Berard
InventorsAlexandre Cafaro, Quentin Spinat, Eric Deutsch, Vincent Gregoire, Nikos Paragios
Life of this patent
Estimating the term…
Abstract
A computer-implemented method for reconstructing a 3D image of a human body part studied of a subject from at least one 2D projection of the human body part studied of the subject and a generative network trained for reconstructing 3D images. The human body part may be the head and/or the neck.
Claims20
A computer-implemented method for reconstructing a 3D image of a human body part studied of a subject from at least one 2D projection of the human body part studied of said subject, said method comprising:
- receiving a first set of latent variables as input data;
- training a generative network based on a library of reference 3D scans of the corresponding human body part to obtain a trained generative model for reconstructing 3D images configured to receive the first set of latent variables as input;
- receiving at least one 2D projection of the human body part studied of the subject;
- obtaining an optimized first set of latent variables by optimizing values of the first set of latent variables by iteratively:
- i) generating with the trained generative model a synthetic 3D image of the studied human body part using the first set of latent variables;
- ii) generating at least one synthetic 2D projection of the synthetic 3D image, the at least one synthetic 2D projection being representative of the at least one received 2D projection of the studied human body part;
- iii) computing an error between the at least one synthetic 2D projection and the at least one received 2D projection of the studied human body part and modifying values of the first set of latent variables for minimizing the error;
- and
- generating a reconstructed 3D image of the studied human body part with the trained generative model using said optimized first set of latent variables.
A device comprising a processor configured to carry out a computer-implemented method for reconstructing a 3D image of a human body part studied of a subject from at least one 2D projection of the human body part studied of said subject, said method comprising:
- receiving a first set of latent variables as input data;
- training a generative network based on a library of reference 3D scans of the corresponding human body part to obtain a trained generative model for reconstructing 3D images configured to receive the first set of latent variables as input;
- receiving at least one 2D projection of the human body part studied of the subject;
- i) obtaining an optimized first set of latent variables by optimizing values of the first set of latent variables by iteratively:
- ii) generating with the trained generative model a synthetic 3D image of the studied human body part using the first set of latent variables;
- iii) generating at least one synthetic 2D projection of the synthetic 3D image, the at least one synthetic 2D projection being representative of the at least one received 2D projection of the studied human body part;
- iv) computing an error between the at least one synthetic 2D projection and the at least one received 2D projection of the studied human body part and modifying values of the first set of latent variables for minimizing the error;
- v) generating a reconstructed 3D image of the studied human body part with the trained generative model using said optimized first set of latent variables (S400).
Dependencies are inferred from the claim text.