An objective evaluation framework for segmentation techniques of functional positron emission tomography studies

Segmentation of multi-dimensional functional positron emission tomography (PET) studies into regions of interest (ROI) exhibiting similar temporal behavior is useful in diagnosis and evaluation of neurological images. Quantitative evaluation plays a crucial role in measuring the segmentation algorithm's performance. Due to the lack of "ground truth" available for evaluating segmentation of clinical images, automated segmentation results are usually compared with manual delineation of structures which is, however, subjective, and is difficult to perform. Alternatively, segmentation of co-registered anatomical images such as magnetic resonance imaging (MRI) can be used as the ground truth to the PET segmentation. However, this is limited to PET studies which have corresponding MRI. In this study, we introduce a framework for the objective and quantitative evaluation of functional PET study segmentation without the need for manual delineation or registration to anatomical images of the patient. The segmentation results are anatomically standardized to a functional brain atlas, where the segmentation of the corresponding MRI reference atlas image is used as the ground truth. We illustrate our evaluation framework by comparing the performance of two pixel-classification techniques based on k-means and fuzzy c-means cluster analysis, applied to clinical dynamic human brain PET studies. The experimental results show that the proposed evaluation framework is able to provide objective measures for segmentation comparison and performance.