Automatic classification of 123I-FP-CIT (DaTSCAN) SPECT images

Abstract

Introduction: We present a method of automatic classification of I-fluoropropyl-carbomethoxy-3β-4-iodophenyltropane (FP-CIT) images. This technique uses singular value decomposition (SVD) to reduce a training set of patient image data into vectors in feature space (D space). The automatic classification techniques use the distribution of the training data in D space to define classification boundaries. Subsequent patients can be mapped into D space, and their classification can be automatically given.

Methods: The technique has been tested using 116 patients for whom the diagnosis of either Parkinsonian syndrome or non-Parkinsonian syndrome has been confirmed from post I-FP-CIT imaging follow-up. The first three components were used to define D space. Two automatic classification tools were used, naïve Bayes (NB) and group prototype. A leave-one-out cross-validation was performed to repeatedly train and test the automatic classification system. Four commercially available systems for the classification were tested using the same clinical database.

Results: The proposed technique combining SVD and NB correctly classified 110 of 116 patients (94.8%), with a sensitivity of 93.7% and specificity of 97.3%. The combination of SVD and an automatic classifier performed as well or better than the commercially available systems.

Conclusion: The combination of data reduction by SVD with automatic classifiers such as NB can provide good diagnostic accuracy and may be a useful adjunct to clinical reporting.