Incorporating Priori Structure Knowledge in Level Set Method-Based Algorithm for 11C-Pittsburgh Compound B Positron Emission Tomography Segmentation

Abstract

Accurate segmentation of 11C-Pittsburgh compound B Positron Emission Tomography (11C-PiB PET) images may assist physicians in diagnosing Alzheimer's disease (AD). However, the weak boundaries of high 11C-PiB concentration regions and the heterogeneous distribution of tracers often bring difficulties for segmentation algorithms. This paper aims at investigating the performance of Level Set Method (LSM)-based algorithms in PiB PET images segmentation. We thereby propose a T-L algorithm that uses the mean-thresholding algorithm (T algorithm) to generate the initial contours for the LSM-based algorithm. Priori structure knowledge, which is derived from the Anatomical Automatic Labeling (AAL) model in corresponding Magnetic Resonance Imaging (MRI) images, is incorporated in the T-algorithm. Seventy-four 11C-PiB PET images were used to evaluate the effectiveness and efficiency of the algorithm. The experiment results indicate that, with respect to the dice similarity coefficient, the Hausdorff distance, and the coefficient of variation values, the proposed T-L algorithm (Dice: 94.8%±2.3%, COV: 0.02; HD:11±11 pixel, COV: 1; computing time: 7.39±0.16s) outperforms both the T-algorithm (Dice: 94.%±2.7%, COV: 0.03; HD: 18±18 pixel, COV: 1; computing time: 67.06±0.14s) and the manual assisted LSM-based algorithm (M-L algorithm) (Dice: 93.4%±3.8%, COV: 0.04; HD: 14±13 pixel, COV: 0.9; computing time: 0.33±0.02s). Furthermore, it is also able to save the time and effort of physicians.