This study presents a methodological framework for AD and MCI diagnosis using radiomic analysis of 18FDG-PET imaging and conducts non-invasive predictions and in-depth analysis of AD and MCI and the associated small number of regions and features. Our methodology follows a structured process commencing with data preprocessing and labeling, facilitating segmentation through FastSurfer, a tool that efficiently segments the brain into 95 ROIs using the DKT-atlas. Subsequently, Feature extraction was carried out using PyRadiomics, calculating 120 features for each of the 95 ROIs (11,400 per image). These extracted features form the foundation of our radiomics analysis, primarily for early diagnostic purposes. In the feature selection phase, we explored a set of eight commonly employed techniques, including ANOVA, PCA, and LASSO, originating from the four main categories, namely filtered, embedded, wrapper, and hybrid methods, to identify a pertinent subset of features. Our evaluation assessed the performance of nine classification methods, such as GradientBoosting, RandomForest, and GaussianNB, in conjunction with eight feature selection techniques. The choice of feature selection method and classifiers was predicated on their ability to achieve the best area under the ROC curve with independent data. For all three predictions AD vs. CN, AD vs. MCI, and CN vs. MCI the Random Forest (RF) classifier with LASSO feature selection demonstrated the highest accuracy with an AUC of 0.976 for AD vs CN, AUC=0.917 for AD vs MCI, and AUC=0.877 for MCI vs CN. In conclusion, our RAB-PET platform enables efficient AD and MCI diagnosis from FDG-PET images using a radiomics pipeline. It also offers a general hardware and software tool for the investigation of other brain disorders.