Linear intensity normalization of DaTSCAN images using Mean Square Error and a model-based clustering approach.

Abstract

The analysis of 3D SPECT brain images requires several pre-processing steps such as intensity normalization and brain feature extraction. In this sense, a new method for intensity normalization of ¹²³I-ioflupane-SPECT (DaTSCAN) brain images based on minimization of the Mean Square Error (MSE) between the Gaussian Mixture Model (GMM)-based extracted features from each subject image and a template in the so-defined non-specific region is derived. Our approach to feature extraction consists of using the set of parameters that define the template features, such as weights, covariance matrices and mean vectors to model the remaining images by reducing, consequently their dimensionality. The proposed method is compared to a widely used approach such as specific-to-non-specific binding ratio normalization. This comparison is performed on a DaTSCAN image database comprising analysis and classification stages for the development of a computer aided diagnosis (CAD) system for Parkinsonian syndrome (PS) detection.