Positioning anatomical landmarks in volume data sets

摘要

Certain embodiments provide a computer apparatus operable to carry out a data processing method to position a set of anatomical landmarks in a three-dimensional image data set of a part or all of a patient, comprising: providing a trained supervised machine learning algorithm which has been trained to place each of the set of anatomical landmarks; applying the supervised machine learning algorithm to place the set of anatomical landmarks relative to the data set; providing a trained point distribution model, including a mean shape and a covariance matrix, wherein the mean shape represents locations of the set of landmarks in a variety of patients; and applying the point distribution model to the set of landmarks with the locations output from the supervised machine learning algorithm by: removing any landmarks whose locations have an uncertainty above a threshold with reference to the mean shape and covariance matrix; followed by: an optimization of the locations of the remaining landmarks by maximizing joint likelihood that a new shape, derived from linear combinations of eigenvectors of the covariance matrix, is plausible.

主权项

1. A computer apparatus to position a set of anatomical landmarks in a three-dimensional image data set of a part or all of a patient, comprising:
processing circuitry configured to provide a trained supervised machine learning algorithm which has been trained to place each of the set of anatomical landmarks; apply the supervised machine learning algorithm to place the set of anatomical landmarks relative to the data set; provide a trained point distribution model, including a mean shape and a covariance matrix, wherein the mean shape represents locations of the set of landmarks in a variety of patients; and apply the point distribution model to the set of landmarks with the locations output from the supervised machine learning algorithm by:
(i) removing any landmarks whose locations have an uncertainty above a threshold with reference to the mean shape and covariance matrix; followed by:(ii) an optimisation of the locations of the remaining landmarks by maximising joint likelihood that a new shape, derived from linear combinations of eigenvectors of the covariance matrix, is plausible.