The Philips/University Medical Center Hamburg-Eppendorf software works by performing three different steps. The first step is to spatially normalize the patient's brain-scan image by selecting, rotating and scaling the appropriate image slice in order to align it with a standard template. The second step is to compare this normalized image, voxel-by-voxel, with a library of normal (un-diseased) brain scans in order to identify hypometabolic regions in the patient's brain. After identifying and color highlighting these hypometabolic regions, the final step is to compare their size, shape and distribution against a set of disease-specific patterns for each type of dementia. The software then quantifies, in the form of a percentage value, the degree to which the patient's scan matches each disease-specific pattern. Ultimately, the diagnosing physician could take the percentages into account when arriving at his or her diagnosis.

Retrospective studies
In addition to being evaluated for feasibility and usability in a clinical setting by University Medical Center Hamburg-Eppendorf's nuclear medicine department, with positive results, the accuracy with which the software can quantify a match between the patient's brain-scan images and disease-specific patterns has been tested in two retrospective studies.
Both of these studies involved using a library of brain scan images, each image having been previously examined by a clinical expert in order to arrive at a differential diagnosis. During the study, each of these images was analyzed by the software and a diagnostic conclusion drawn from its results, based on the disease-specific match percentages generated. These diagnostic conclusions were then compared to the differential diagnoses made previously by the clinical expert. Both studies employed a so-called 'leave-one-out cross-validation scheme', in which each patient's brain scan (the validation data) was compared to the disease-specific patterns in all the other brain scans (the training data). This is a well-known scheme for minimizing bias in tests where the data set size is limited.

In the first study, based on a University Medical Center Hamburg-Eppendorf library of FDG-PET scans from 83 patients, the software achieved better than 98% correspondence with the clinical expert's interpretation, when programmed to differentiate between brain scans showing no signs of dementia, those showing characteristics of Alzheimer's disease and those showing characteristics of Frontotemporal Dementia.

More Industry Headlines