Method: Currently, many challenges in segmentation, feature design and modelling make medical image mining a labour intensive process that requires medical expertise. Consequently, much of the information in medical image databases is currently not effectively used to support diagnosis, research and education. We propose to sidestep segmentation and feature design by automatically extracting general purpose, localized visual features using Gabor filters. We then sidestep model construction and model based classification by applying an ensemble of of case based image reconstruction methods that yield a sparse presentation of the new image. This combination of techniques offers an easy to deploy system for retrieving useful old images that are similar to the new image from image databases.
The symbolic information in the old images can then be used to automatically suggest annotations to the new image. In our work we have used the pathology class information attached to the computer tomography (CT) images of the traumatic brain injury (TBI) to suggest classification of the new images. However, the method is not specific to CT scans and it scales well to large image databases.
Results: We applied our method to 847 CT images of TBI obtained from the database of the Neuroradiology Department in a tertiary referral hospital specializing in neurological diseases in Singapore. Our stratified cross-validation results demonstrate the capability of our method to automatically classify the types of traumatic brain injuries into subdural hematoma, extradural hematoma, and intracerebral hemorrage. This functionality allows searching for medical images by their diagnosis based on the image content only. We also demonstrate a tool that shows the relevant images used in these automatic classifications.
Conclusions: Our method offers an easy way to use information in medical image databases. The tool based on the methodology can be used to support diagnosis, and possibly in future, prognosis in medical decision making process.