Image-based tumor phenotyping with machine learning from synthetic data
Machine training and application of machine-trained classifier are used for image-based tumor phenotyping in a medical system. To create a training database with known phenotype information, synthetic medical images are created. A computational tumor model creates various examples of tumors in tissue. Using the computational tumor model allows one to create examples not available from actual patients, increasing the number and variance of examples used for machine-learning to predict tumor phenotype. A model of an imaging system generates synthetic images from the examples. The machine-trained classifier is applied to images from actual patients to predict tumor phenotype for that patient based on the knowledge learned from the synthetic images.
1. A method for image-based tumor phenotyping in a medical system, the method comprising:
acquiring medical scan data representing a tumor in tissue of a patient;
extracting a set of features from the medical scan data;
inputting the features to a machine-trained classifier, the machine trained classifier trained, at least in part, from training data comprising images generated synthetically from tumor models to not be specific to existing patient cases and having different tumor pathologies than in a database of the training data, the images generated synthetically based on a tumor model;
determining, by the medical system using application of the machine-trained classifier to the features, phenotype information of the tumor; and
outputting the phenotype information characterizing the tumor in the tissue of the patient.
2. The method of claim 1 wherein acquiring comprises acquiring computed tomography data, magnetic resonance data, or molecular imaging data.
3. The method of claim 1 further comprising obtaining other data for the patient, the other data comprising symptom, patient demographics, blood biomarkers, patient history, non-invasive measurements, extracts from a clinical report, genetics, or combinations thereof, wherein inputting comprises inputting the other data, and wherein determining comprises determining by application of the machine-trained classifier to the features and other data.
4. The method of claim 1 wherein the machine-trained classifier comprises a deep learnt classifier, and wherein extracting comprises extracting the set of features as part of the determining where the application inputs the features to a subsequent layer of the machine-trained classifier.
5. The method of claim 1 wherein determining comprises determining with the application of the machine-trained classifier where the machine-trained classifier is trained, at least in part, from the synthetically generated images, where the synthetically generated images are a combination of a model of one or more scan modality with the tumor model.
6. The method of claim 1 wherein determining comprises determining with the application of the machine-trained classifier where the machine-trained classifier is trained, at least in part, from the synthetically generated images, where the synthetically generated images are generated using a multi-scale model of tumor physiology as the tumor model, the multi-scale model including molecular, cellular and tissue scales.
7. The method of claim 1 wherein determining comprises determining with the application of the machine-trained classifier where the machine-trained classifier is trained, at least in part, from the synthetically generated images, where the synthetically generated images are generated from perturbation of the tumor model.
8. The method of claim 1 wherein determining comprises determining with the application of the machine-trained classifier where the machine-trained classifier is trained, at least in part, from the synthetically generated images, where the synthetically generated images are generated from a machine-learnt model of cellular physiology of the tumor model.
9. The method of claim 1 wherein determining comprises determining with the application of the machine-trained classifier where the machine-trained classifier is trained, at least in part, from the synthetically generated images, where the synthetically generated images are generated with at least one model of a scanning modality.
10. The method of claim 1 wherein determining the phenotype information comprises determining a diagnosis of the tumor and/or a therapy response.
11. A method for machine training for image-based tumor phenotyping in a medical system, the method comprising:
varying parameters via interactively or automatically of a computational tumor model, the varying providing a set of synthetically generated examples of tumors including examples not specific to any actual patients and having different virtual pathologies than in any training database;
emulating, by simulation of imaging, medical images from the synthetically generated examples of tumors; and
training, with a machine of the medical system, a machine-learnt classifier using the emulated medical images as training data with samples including the different virtual pathologies.
12. The method of claim 11 wherein varying comprises varying the parameters with the computational tumor model comprising a multi-scale model, a machine-learnt model, or both.
13. The method of claim 12 wherein the machine-learnt model comprises a machine learnt model of cellular state.
14. The method of claim 11 wherein varying comprises varying with the computational tumor model including interaction of the tumors with healthy tissue, and wherein emulating comprises emulating from the examples of the tumors and healthy tissue.
15. The method of claim 11 wherein emulating comprises generating parametric maps of characteristics that interact with an imaging modality from the examples of the tumors, and simulating a scan by the imaging modality using the parametric map.
16. The method of claim 11 wherein training comprises training the machine-learnt classifier to predict a tumor grade, tumor stage, therapy response, benign or malignant, tumor cell differentiation, or prognosis.
17. The method of claim 11 wherein training comprise training a deep-learnt classifier as the machine-learnt classifier.
18. The method of claim 11 wherein varying comprises varying an amount of time, and wherein training comprises training the machine-learnt classifier to classify from patient images over time.
19. A method for machine training for image-based tumor phenotyping in a medical system, the method comprising:
modeling tumor physiology via interactively or automatically changes, the modeling providing samples of various tumor anatomies different than in a training database;
creating synthetic images with an image model from the various tumor anatomies from the modeling of tumor physiology; and
training a machine by the machine training to predict from just the database of training data including the images created from the various tumor anatomies from the modeling of the tumor physiology.
20. The method of claim 19 further comprising applying the trained machine to an image of a patient to predict a tumor characteristic of a tumor represented in the image of the patient.