Knowledge-based automatic image segmentation

A method, comprising:

• accessing, from computer system memory, a medical image set for a patient, wherein the medical image set comprises a plurality of structures of interest of the patient;
• selecting clusters of pixels in the medical image set that are statistically probable to be inside the structures of interest according to a statistical model comprising weighted-average segmentations of the structures of interest to generate sparse landmark signatures for the medical image set;
• selecting a plurality of reference image sets from a knowledge-base to form a base set, wherein sparse landmark signatures of the selected plurality of reference image sets match the sparse landmark signatures of the medical image set in the base set;
• determining a first weighted-average segmentation using each structure of interest in the base set;
• searching the knowledge-base and identifying matches to the plurality of points to form a plurality of sub-volume base sets, wherein a sub-volume base set comprises a plurality of reference image set sub-volumes;
• determining a second weighted-average segmentation using each structure of interest in the plurality of structures of interest in the medical image set;
• and
• updating the statistical model for generating sparse landmark signatures with results from the second weighted-average segmentation.

A non-transitory computer readable storage media having computer-readable and computer-executable program code embodied therein for causing a computer system to execute a method, the method comprising:

• accessing a medical image set for a patient, wherein the medical image set comprises a structure of interest of the patient;
• selecting a cluster of pixels in the medical image set that are statistically probable to be inside the structure of interest according to a statistical model comprising weighted-average segmentations of the structure of interest to generate sparse landmark signatures for the medical image set;
• selecting a plurality of reference image set from a knowledge-base to form a base set wherein sparse landmark signatures of the selected plurality of, reference image sets match the sparse landmark signatures of the medical image set in the base set, and wherein the base set is formed from a cohort of image sets within the knowledge-base that share common characteristics, wherein the common characteristics comprise at least one of:
• sex, diagnosis, stage, responder versus non-responders, age, treatment site, and size;
• determining a weighted-average segmentation based on a statistical weighted average of deformation of the base set to the medical image set;
• and
• updating the statistical model for generating sparse landmark signatures with results from the weighted-average segmentation.