Computer Aided Alzheimer’s Diagnosis Based On 11C-PiB PET Images

Alzheimer’s disease(AD) is an irreversible neurodegenerative disease. However, the cause for most Alzheimer’s cases is still mostly unknown. The amyloid hypothesis postulated that extracellular amyloid beta(Aβ) deposits in the brains of AD patients are one of the fundamental cause of the disease. Hence, diagnosing AD in the early stage is significant. Recently, the tracer 11C-labeled Pittsburgh Compound-B(11C-Pi B) specially binds to amyloid-beta(Aβ) fibers and plaques, thus enabling the in vivo detection of this hallmark of AD, and it contributes to AD classification and rehabilitation assessment as well. Acutually, diagnosing AD by 11C-Pi B Positron Emission Tomography(PET) images is primarily based on visual assessment, which rely on subjective experience of radiologists, leading to inaccurate and inadequate results. Thus, an accurate Computed Aided AD Diagnosis(CAAD) is valuable for clinically use.This study therefore proposes a CAAD approach for radiologists to analyse 11 CPi B PET image. The proposed CAAD includes four parts: image pre-processing; 3D thresholding-Lattice Boltzmann Method(3D T-LB) based algorithm is for fast Aβ regions of interest(ROIs) segmentation; Principal Component Analysis(PCA) and Recursive feature elimination using the support vector machine(SVM) weight map(RFE W-map) are used for feature extraction and selection respectively; SVM is for classification.In order to verfy the accuracy and clinical significance of the proposed CAAD, four comparison experiments were carried out by selecting a total of 149 11C-Pi B PET data from ADNI database and PET center of Huashan hospital(Shanghai). In preprocessing, the visual assessment experiments were conducted by the radiologists and professional graduates to test the image quality. In ROIs segmentation, the golden standard was drawn by clinical radiologists, and 3D T-LB algorithm was compared with other two PET image segmentation algorithms. In feature extraction and selection, a statistical experiment between these features and Mini-Mental State Examination(MMSE) scores was set up, and the classification experiments between different feature selection methods were conducted as well. In classification, three classifiers were compared to test the classification performance of the CAAD in differentiating AD, mild cognition impairment(MCI) and healthy control(HC) groups, and cross validation was achieved by means of the leave-one-out method.The experimental results showed that the novel CAAD can provide a good effect for AD diagnosis in all four parts. In pre-processing, the results kept in consistent with visual assessments. In ROIs segmentation, 3D T-LB algorithm was obviously prior to other algorithms in effectivity and efficiency, and it could realize an 88% accuracy at least while comparing with the golden standard. In feature extraction and selection, most features could meet the statistical significance, and results showed a good correlation between features and MMSE scores. Moreover, RFE W-map was supposed to be the best feature selection method while compared with others in classification performance. In classification, under the SVM with polynomial kernel, the proposed CAAD performed best in differentiating AD, MCI and HC groups. As a result, the proposed CAAD could yielid a 95.97% accuracy in AD diagnosis, and it was also prior to the existing CAADs from literatures(including visual assessment and CAAD approaches for other images). Furthermore, the efficiency of the proposed CAAD approach was accepted in clinics.