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

Area of Science:

• Medical Imaging
• Neuroscience
• Computer-Aided Diagnosis

Background:

• 3D SPECT brain image analysis requires pre-processing like intensity normalization and feature extraction.
• Accurate normalization is crucial for reliable analysis and the development of diagnostic systems.

Purpose of the Study:

• To introduce a new intensity normalization method for 123I-ioflupane-SPECT (DaTSCAN) brain images.
• To evaluate the proposed method's effectiveness in feature extraction and dimensionality reduction.
• To compare the new method against established normalization techniques for Parkinsonian syndrome detection.

Main Methods:

• A novel intensity normalization technique based on minimizing Mean Square Error (MSE) between Gaussian Mixture Model (GMM) features of subject images and a template non-specific region.
• Feature extraction using GMM parameters (weights, covariance matrices, mean vectors) for dimensionality reduction.
• Comparison with specific-to-non-specific binding ratio normalization using a DaTSCAN image database.

Main Results:

• The proposed GMM-based normalization method effectively extracts features and reduces dimensionality.
• The method demonstrates potential for improving the analysis and classification stages in computer-aided diagnosis (CAD) systems.
• Performance evaluation on a DaTSCAN database for Parkinsonian syndrome detection.

Conclusions:

• The developed GMM-based intensity normalization method offers a promising approach for pre-processing DaTSCAN images.
• This technique can enhance the accuracy of feature extraction and dimensionality reduction in SPECT brain imaging.
• The method shows potential for advancing CAD systems in the detection of Parkinsonian syndrome.