Query by Image of Brain SPECT Database

Abstract

In contrast to search engines for medical image databases that use age, gender, or disease classification, Query by Image search allows an image to be presented as input to the search engine. Images together with clinical reports are returned that best match the presenting image. Work on query by image systems have been ongoing for more than two decades, predominately in fields outside of medical imaging. In these fields, features are often identified and used in the search for similar images. Corresponding strategies have been taken in medical imaging, especially MRI where it is reasonable to match based on a feature. However, in brain SPECT imaging, clinicians are often interested in the global pattern of brain activity for making diagnoses such as Small Vessel Disease, Mild Cognitive Impairment, Parkinson’s or Alzheimer’s Disease, which have some commonalities as “Global Brain Impairment” patterns. By utilizing robust spatial normalization methods to transform images to a common stereo-tactic space, we are able to use simple methods for measuring and ranking the closeness between the presenting image and images in the database. Our decomposition of the Brain SPECT dataset shows that images within the dataset have very high similarity. However, subtle differences can be reliably utilized for selecting best image matches. Throughout testing, highly relevant cases were consistently returned for our image queries. Our method is fast, robust, intuitive for users, and practical.