Final Report Summary - MEDYMA (Biophysical Modeling and Analysis of Dynamic Medical Images.)
In the MedYMA project we achieved significant progress towards the original three main goals:
1) design of a new generation of biophysical models of organs or pathologies (heart, brain,etc.),
2) construction of databases of synthetic but realistic dynamic medical images,
3) automatic personalization of biophysical models from the patient's medical images.
We introduced those biophysical models not only to quantify the evolution of pathology from medical images, but also to simulate time-series of images with prescribed values of some biophysical parameters. We used the simulated images with known values of the biophysical parameters to better understand the development of the disease and also as input for machine learning algorithms.
In the cardiac domain, we developed a new generation of biophysical models of the beating heart. These models take into account the electrical and the mechanical activity of the myocardium with a limited number of key parameters. We showed that these models could be personalized from medical images and biological signals. We showed for the first time the possibility to automatically calibrate and then identify a number of regional biophysical parameters of the beating heart. We also demonstrated for the first time the potential use of these parameters to characterize the condition of a person on a database of patients and controls. We also used our cardiac biophysical models to generate synthetic but realistic cardiac images (MR, CT and Ultrasound), and we used these synthetic images with known biophysical parameters to train machine-learning algorithms and learn how to automatically detect the propagation of the electrical wave in clinical images, or to detect and locate the presence of ischemic tissues affecting the cardiac motion.
For brain tumors, we developed a new generation of biophysical models to describe the proliferation and migration of tumoral cells. We showed that these models could be personalized from a time series of brain MR images, and/or by the analysis of the irregularity of the front of the tumor in a single MR image. We used our brain tumor models to generate synthetic but realistic MR images (T2, T1, T1 with gadolinium and Flair sequences) and showed how a machine learning algorithm could use those sequences along with the parameters of the biophysical model to learn the tumoral density, first on synthetic sequences, and then on clinical images. We simulated brain tumor evolutions with various diffusion parameters to better understand the role of white matter fibers in the anisotropic migration of tumoral cells. We also coupled our biophysical model of brain tumors with a biophysical model of radiotherapy in order to optimize radiotherapy protocols. We also proposed a probabilistic patch-based model of brain tumor appearance to analyze MR images of brain tumors and we used the same framework to simulate synthetic but realistic MR images of brain tumors.
For Alzheimer's disease, we developed a new generation of geometrical and biophysical models of brain atrophy. The first class of models is based on an advanced diffeomorphic registration of time series of brain MRIs. The LCC-Demons software is freely available on our web site. A second class of models is based on a biomechanical model of local atrophy, for which a loss of matter can be prescribed in each volume element, inducing a global biomechanical deformation of the brain parenchyma. We showed for the first time the possibility to simulate a spatially varying prescribed atrophy in the brain parenchyma. We also used our geometrical model of brain atrophy to extrapolate the evolution measured during a limited time interval to a longer time interval in the past and in the future. We showed how to perform statistical measures of brain evolution on a population of subjects by transporting velocity fields in a common brain reference. We used this approach to compare various populations of controls and Alzheimer’s disease patients in order to show significant statistical differences in localized regions of the brain. We were also able to propose an average model of normal atrophy, and to decompose a pathological atrophy into a normal and an abnormal component. Finally, we used the biophysical model of brain atrophy to simulate realistic time-series of MR images of brains of virtual patients with neurodegenerative diseases for which we can control at each point the atrophy rate. The Simul@trophy software is freely available on our web site.
1) design of a new generation of biophysical models of organs or pathologies (heart, brain,etc.),
2) construction of databases of synthetic but realistic dynamic medical images,
3) automatic personalization of biophysical models from the patient's medical images.
We introduced those biophysical models not only to quantify the evolution of pathology from medical images, but also to simulate time-series of images with prescribed values of some biophysical parameters. We used the simulated images with known values of the biophysical parameters to better understand the development of the disease and also as input for machine learning algorithms.
In the cardiac domain, we developed a new generation of biophysical models of the beating heart. These models take into account the electrical and the mechanical activity of the myocardium with a limited number of key parameters. We showed that these models could be personalized from medical images and biological signals. We showed for the first time the possibility to automatically calibrate and then identify a number of regional biophysical parameters of the beating heart. We also demonstrated for the first time the potential use of these parameters to characterize the condition of a person on a database of patients and controls. We also used our cardiac biophysical models to generate synthetic but realistic cardiac images (MR, CT and Ultrasound), and we used these synthetic images with known biophysical parameters to train machine-learning algorithms and learn how to automatically detect the propagation of the electrical wave in clinical images, or to detect and locate the presence of ischemic tissues affecting the cardiac motion.
For brain tumors, we developed a new generation of biophysical models to describe the proliferation and migration of tumoral cells. We showed that these models could be personalized from a time series of brain MR images, and/or by the analysis of the irregularity of the front of the tumor in a single MR image. We used our brain tumor models to generate synthetic but realistic MR images (T2, T1, T1 with gadolinium and Flair sequences) and showed how a machine learning algorithm could use those sequences along with the parameters of the biophysical model to learn the tumoral density, first on synthetic sequences, and then on clinical images. We simulated brain tumor evolutions with various diffusion parameters to better understand the role of white matter fibers in the anisotropic migration of tumoral cells. We also coupled our biophysical model of brain tumors with a biophysical model of radiotherapy in order to optimize radiotherapy protocols. We also proposed a probabilistic patch-based model of brain tumor appearance to analyze MR images of brain tumors and we used the same framework to simulate synthetic but realistic MR images of brain tumors.
For Alzheimer's disease, we developed a new generation of geometrical and biophysical models of brain atrophy. The first class of models is based on an advanced diffeomorphic registration of time series of brain MRIs. The LCC-Demons software is freely available on our web site. A second class of models is based on a biomechanical model of local atrophy, for which a loss of matter can be prescribed in each volume element, inducing a global biomechanical deformation of the brain parenchyma. We showed for the first time the possibility to simulate a spatially varying prescribed atrophy in the brain parenchyma. We also used our geometrical model of brain atrophy to extrapolate the evolution measured during a limited time interval to a longer time interval in the past and in the future. We showed how to perform statistical measures of brain evolution on a population of subjects by transporting velocity fields in a common brain reference. We used this approach to compare various populations of controls and Alzheimer’s disease patients in order to show significant statistical differences in localized regions of the brain. We were also able to propose an average model of normal atrophy, and to decompose a pathological atrophy into a normal and an abnormal component. Finally, we used the biophysical model of brain atrophy to simulate realistic time-series of MR images of brains of virtual patients with neurodegenerative diseases for which we can control at each point the atrophy rate. The Simul@trophy software is freely available on our web site.