Comparative Study of Segmentation Techniques for Basal Ganglia Detection Based on Positron Emission Tomography Images

The Basal Ganglia (BG) is a small brain structure that plays a significant role in the Parkinson’s disease (PD) pathogenesis. Since it provides detailed quantitative information for the PD progression, positron emission tomography (PET) is an ideal tool for the detection of the amount changes in this region. To make these images easy to analyze and more meaningful, it is crucial to modify their representation and clarify the image through segmentation. It is critical to locate the BG region so that it can be easily examined to detect its size and shape for progression diagnosis. In this paper, an exhaustive study of BG segmentation from PET images is performed through different segmentation techniques, namely binary thresholding, truncate thresholding, threshold to zero, Otsu thresholding, and clustering combined with binary thresholding. These techniques have been tested onto 110 PET images and evaluated with their corresponding ground truth which were previously segmented manually. We evaluated the segmentation performance using two metrics Dice Similarity Coefficient (DSC) and mean Intersection over Union (mIoU). The obtained results indicate that binary thresholding technique outperforms other segmentation techniques and reached higher performance using 150 as threshold with DSC of 0.7701 and mIoU of 0.6394.