Probabilistic Refinement for Model-Based Segmentation

Abstract

A system for segmenting current diagnostic images includes a workstation (30) which segments a volume of interest in previously generated diagnostic images of a selected volume of interest generated from a plurality of patients. One or more processors (32) are programmed to register the segmented previously generated images and merge the segmented previously generated images into a probability map that depicts a probability that each voxel represents the volume of interest (24) or background (26) and a mean segmentation boundary (40). A segmentation processor (50) registers the probability map with a current diagnostic image (14) to generate a transformed probability map (62). A previously-trained classifier (70) classifies voxels of the diagnostic image with a probability that each voxel depicts the volume of interest or the background. A merge processor (80) merges the probabilities from the classifier and the transformed probability map. A segmentation boundary processor (84) determines the segmentation boundary for the volume of interest based on the current image based on the merge probabilities.

Description

technical field [0001] This application deals with image segmentation. It can especially be applied in conjunction with medical diagnostic imaging for delineating target volumes, organs, etc. Background technique [0002] Various diagnostic imaging modalities such as CT, MRI, PET, SPECT, and ultrasound generate three-dimensional images of a patient's internal anatomy. Different organs, different tissues, cancerous tissue in contrast to non-cancerous tissue, etc. are typically depicted with different gray levels, which can be mapped into different colors for easier differentiation. Adjacent organs, tissue volumes, etc. often have little or no significant grayscale difference. For example, some soft tissues may have poor contrast in CT data. Such poor or blurred contrast makes the corresponding border part only partially visible, ie unclear, not clearly defined. [0003] This problem has been addressed with model-based segmentation. Typically, some areas of the boundary a...

AI Technical Summary

Problems solved by technology

[0005] Unfortunately, imposing constraints on the model shape may not be conducive to accurately following the boundaries of the structure or organ under study

Finding the optimal balance between the two energy terms is often not an easy task and can lead to ambiguous solutions or multiple potential solutions

Images