Random Forest and K-Means Clustering Algorithms to Classify of 18F-Florbetapir Brain PET

This paper explores and compares the use of two common machine learning (ML) algorithms, random forests (RF) and k-means clustering (KMC), for classifying 18F-florbetapir brain PET as positive or negative for amyloid deposition. The pilot dataset consists of 65 18F-Florbetapir PET and corresponding MRI studies taken from the Alzheimer's Disease Neuroimaging Initiative (ADNI), in patients with mild cognitive impairment (MCI). Each PET scan was read as positive or negative for amyloid deposition by two physicians dual board certified in nuclear medicine and radiology with final interpretation based on consensus. This clinical interpretation of the PET scans served as the gold standard. Using an image processing pipeline, standardized uptake value ratios (SUVR) were computed in 57 brain regions, with normalization to the cerebellar gray matter. The RF algorithm had a slightly higher classification accuracy (91±6%) compared with the KMC algorithm (81±3%), using 4-fold cross-validation. However, the KMC algorithm had lower computational cost and may highlight equivocal cases on clinical interpretation. Further investigation is ongoing.