ROI-based assessment of abnormality using transformation invariant features

摘要

A method of comparing the results of medical imaging by (e.g.) Positron Emission Tomography (PET) scanning is disclosed which dispenses with the need for intensity normalization. The relationships between features extracted from relevant regions of interest in the image are studied. In one example, mean intensities in the principle brain lobes are compared to each other and a short image ID is constructed and used to derive population statistics and diagnosis. A set of regions of interest (ROIs) may be defined in a test image and a corresponding set of ROIs in each of a set of reference images. For each ROI in each image the value of a chosen linear feature (eg average intensity) is calculated. For each image, a set of ratios is calculated by dividing the value of the linear feature in an ROI with the value of the linear feature in another ROI within the same image. Statistical analysis of the distribution of ratios (eg Z-scores) is used to compare the test image to the reference images. The population statistics can be compared with 'reference' statistics in order to assess abnormality. Comparison by a number of methods is possible and a further feature of the invention concerns a novel voting mechanism which derives abnormality scores for each region. The system may be used to detect Alzheimer's disease.