Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11847786-B2
Publication Date
2023-12-19
Expiration Date
Abstract
A machine learning model is described that is trained without labels to predict a motion field between a pair of images. The trained model can be applied to a distinguished pair of images to predict a motion field between the distinguished pair of images.
Core Innovation
A method in a computing system trains a machine learning model to predict the value of a motion field between a pair of radiological images captured from the same patient at different times. For each training pair, the method applies the machine learning model to first and second radiological images to obtain a predicted motion field value, applies a spatial transform to transform the first radiological image using the predicted motion field value, and compares the transformed first radiological image to the second radiological image to obtain a loss metric value.
The method further determines an inverse of the predicted motion field value and applies a spatial transform to transform the second radiological image using the determined inverse. It compares the transformed second radiological image to the first radiological image to obtain an additional loss metric value, and adjusts the state of the machine learning model based on a sum of the loss metric value and the additional loss metric value. The method stores the trained state of the machine learning model.
A memory-based representation is also provided in one or more computer memories not constituting a carrier wave, collectively containing a machine learning model data structure. The data structure comprises information representing a machine learning model trained without labels to predict a motion field between a pair of images, where the representation is trained by determining an inverse of the predicted motion field value, applying spatial transforms, comparing images to obtain loss metric values, and adjusting the model state based on the sum. The contents are usable to instantiate the trained model for application to a distinguished pair of images to predict a motion field.
Claims Coverage
The partial content includes four independent claims (clm-00001, clm-00014, clm-00021, and clm-00026). Across these claims, the inventive features focus on label-free training of a motion-field predictor from same-patient radiological image pairs at different times, image similarity losses driven by spatial transforms, inverse-consistency via an inverse motion field, region-of-interest segmentation mask guided contour transform and contour loss, and using the trained model to determine a strain result for a distinguished patient.
Inverse motion field consistency training with dual loss metrics
Training the machine learning model by predicting a motion field between a pair of radiological images, applying spatial transforms to transform both images using the predicted motion field value and an inverse of the predicted motion field value, comparing the transformed images to obtain a loss metric value and an additional loss metric value, and adjusting the model based on a sum of the loss metric value and the additional loss metric value.
Label-free trained motion-field model data structure for distinguished pair inference
Storing, in one or more computer memories, a machine learning model data structure in which the model is trained without labels to predict a motion field between a pair of images using an inverse of the predicted motion field value, spatial transforms, comparisons to obtain a loss metric value and an additional loss metric value, and model-state adjustment based on the sum; such that the contents are usable to instantiate the trained model to predict a motion field between a distinguished pair of images.
Region of interest segmentation mask with contour transform and contour loss
Determining a region of interest segmentation mask against training pairs of radiological images, applying a contour transform to transform the first radiological image using the predicted motion field value and the region of interest segmentation mask, comparing the contour-transformed first radiological image to the second radiological image to obtain a contour loss metric value, and adjusting the model state based on the loss metric value and the contour loss metric value.
Predict motion field for production pair and determine strain result
Receiving a production pair of radiological images captured from a distinguished patient at different times, applying the accessed trained machine learning model to produce a predicted motion field for the production pair, and determining a strain result for the distinguished patient using the predicted motion field for the production pair of radiological images.
Overall, the claims cover unsupervised training of a motion-field prediction model from same-patient radiological image pairs at different times using spatial transforms and loss metrics, enhanced by inverse consistency via an inverse motion field and/or by region-of-interest segmentation mask guided contour transform and contour loss, and then using the trained model to predict motion fields and determine a strain result for a distinguished patient.
Stated Advantages
Not explicitly described in patent.
Documented Applications
Not explicitly described in patent.
Interested in licensing this patent?