Granted Patent B2

US 12,322,510 · App. 18/175,337 · Granted Jun 3, 2025

Inventors: Igor Nobrega (Tampa, FL); Wenbin Mao (Tampa, FL)

Assignee: UNIVERSITY OF SOUTH FLORIDA

G16H50/20

Abstract

Claims (37)

1. A method for providing active contraction properties of a myocardium, the method comprising:

inputting a plurality of clinical metrics into a deep learning model;

inputting a representation of a cardiac cycle through a pressure-volume loop into the deep learning model, the deep learning model including a first process layer with a first intermediate output and a second process layer that receives the first intermediate output as a first intermediate input; and

outputting one or more contraction properties of the myocardium.

2. The method of claim 1 , wherein the clinical metrics are extracted from one or more medical imaging modalities.

3. The method of claim 1 , wherein the clinical metrics include longitudinal shortening, radial shortening, wall thickening, longitudinal strain, circumferential strain, and ejection fraction.

4. The method of claim 1 , wherein at least one contraction property includes fiber orientation.

5. The method of claim 1 , wherein at least one contraction property includes a gamma waveform.

6. The method of claim 1 , wherein the first process layer includes a first stage, and

wherein the first stage receives the pressure-volume loop and treats the pressure-volume loop as a static signal with two channels and reduces the data dimension by employing a series of convolution neural network encoders to produce PV codes that represent characteristics of the pressure-volume loop.

7. The method of claim 6 , wherein the first process layer includes a second stage, and

wherein the second stage concatenates the PV codes with the clinical metrics to provide concatenated data.

8. The method of claim 7 , wherein the first process layer includes a third stage,

wherein the third stage takes the concatenated data and employs a succession of fully connected layers with local skip connections at every two layers and general skip connections to initial information at every five layers to generate a set of fiber orientations,

and wherein the set of fiber orientations are configured as the first intermediate output.

9. The method of claim 1 , wherein the second process layer includes a first stage, and

wherein the first stage concatenates the pressure volume-loop with the clinical metrics and the first intermediate output.

10. The method of claim 9 , wherein the second process layer includes a second stage, and

wherein the second stage employs a recurrent neural network to retrieve gamma waveform as a time-distributed sequence.

11. The method of claim 1 , wherein the clinical metrics are estimated from a machine learning model that has been pre-trained on a time-independent dataset of geometric characteristics.

12. The method of claim 1 , wherein the pressure-volume loop is provided by a lumped parameter model that generates a broad spectrum of physiological and pathological pressure-volume loops.

13. The method of claim 1 , wherein one or more of the clinical metrics or the representation of a cardiac cycle includes patient-specific data.

14. A method of data generation for training, validating, or testing a model for outputting properties of a myocardium, the method comprising:

inputting an initial pressure-volume loop into a training model;

inputting initial fiber orientations into the training model;

combining pressure and volume values with fiber orientations in a first intermediate model of the training model to produce geometric characteristics to form synthetic clinical metrics;

combining pressure and volume values with fiber orientations in a second intermediate model of the training model to generate a synthetic gamma waveform; and

supplying the pressure-volume loop and the synthetic clinical metrics to a deep learning model of the training model.

15. The method of claim 14 , further comprising:

outputting, from the deep learning model, an estimated gamma waveform based on the pressure-volume loop and synthetic clinical metrics.

16. The method of claim 14 , wherein the initial pressure-volume loop is generated pseudo-randomly.

17. The method of claim 14 , wherein the pressure and volume values are taken at an end-diastolic period and an end-systolic period.

18. The method of claim 14 , wherein forming the synthetic clinical metrics and generating the gamma waveform are executed in parallel.

19. A method of using a constitutive model in a clinical application when limited clinical data is available, the method comprising:

extracting constitutive model parameters from basic clinical measures, the basic clinical measures including a pressure-volume loop and measurements in only two timesteps of a left ventricle,

the constitutive model configured to correlate applied forces with a material's mechanical response by incorporating characteristics of morphology of tissue into its internal composition to provide tissue behavior analysis.

20. The method of claim 19 , wherein one of the two timesteps is at an end-systole time frame and the other of the two timesteps is at an end-systole time frame.

Continuity (2)

References Cited (116)