DOI: 10.1002/alz.080326 ISSN: 1552-5260

A radiomic features study in 18F‐FDG‐PET images to classify Alzheimer Disease patients using machine learning algorithms

Moisés Ebenezer Hernández Cruz, Arturo Avendaño Estrada, Miguel Ángel Ávila Rodríguez
  • Psychiatry and Mental health
  • Cellular and Molecular Neuroscience
  • Geriatrics and Gerontology
  • Neurology (clinical)
  • Developmental Neuroscience
  • Health Policy
  • Epidemiology

Abstract

Background

Radiomic features has demonstrate their utility to obtain important information in many diseases, specifically information about tissues or their metabolism that can be related to the diagnostic of a patient. In study of the Alzheimer Disease, there are some works in the use of radiomic features measured in Positron Emission Tomography (PET) images with [18F]fluorodeoxyglucose ([18F]FDG) to train machine learning algorithms and classify patients with and within the disease.

Method

A sample of 50 [18F]FDG‐PET studies of patients with neurological disease suspected was used. All PET images were normalized to a brain atlas space and then radiomic features were measured in all volumes of these atlas. Then the most important features correlated with the diagnosis of the Alzheimer Disease were found and used to train different machine learning algorithms and predict the diagnostic of patients. The shipping diagnosis was used as the target for training, and the same process using the classification of an Alzheimer Disease software tool called PMOD as the target was made.

Result

Some features that shows a correlation with the shipping diagnosis of patients and with the classification of the Alzheimer software tool PMOD were found. Most of them are features measured in parietal regions, and from a radiomic features family called Grey Level Co‐ocurrence Matrix. These features are related with tissue texture. Features with the most significant correlation with their shipping diagnosis and with PMOD classification were used to train machine learning algorithms and good results were obtained for both cases, with maximum accuracies of 0.91.

Conclusion

The correlation analysis helped to improve the training process, obtaining acceptable performance results in the classification of patients using their shipping diagnosis and PMOD classifications as the target. Some of these features shows a clear difference in their values between Alzheimer and control patients, perhaps a bigger sample of patients is needed to be sure of the relation between the feature and the disease. The results in this study suggest that it is possible to find radiomic features that can help us in understanding and diagnosing of Alzheimer Disease.

More from our Archive