A computational model to describe multi-regional brain architecture during neurodegeneration in Alzheimer's disease.
Authors
Affiliations (4)
Affiliations (4)
- Department of Surgery and Cancer, Imperial College Hammersmith Campus, Du Cane Road, London, W12 0NN, UK.
- The Imaging Department, Imperial College Healthcare NHS Trust, UK Hammersmith Hospital, Du Cane Road, London, W12 0HS, UK.
- Imperial College Memory Research Centre, Department of Brain Science, Imperial College Healthcare NHS Trust, UK Hammersmith Hospital, Du Cane Road, London, W12 0HS, UK.
- Department of Surgery and Cancer, Imperial College Hammersmith Campus, Du Cane Road, London, W12 0NN, UK. [email protected].
Abstract
We previously proposed an MRI-based machine learning model to describe the mesoscopic architecture of the human brain to aid in classifying subjects as having non-AD related pathology (nADrp) or AD related pathology (ADrp), including mild cognitive impairment (MCI) and Alzheimer's disease (AD). The method, developed on data from patients scanned at 1.5T showed high performance, but did not generalise well to scans obtained from 3T MRI. In the current work we overcome the problem and extend the approach to patients scanned longitudinally, and at different field strengths. Retrospective T1-MRI data from 1592 subjects scanned at 3T were included to develop the machine learning models. Three additional longitudinal datasets (n = 211) at different magnetic field strengths-1.5 and 3T-were adopted to evaluate the models. Radiomic features were extracted from each brain region. A logistic regression method with least absolute shrinkage and selection operator (LASSO) model selection was employed to classify nADrp from ADrp (classifier 1) or MCI from AD (classifier 2). Classifier 1 that discriminates nADrp from ADrp achieves high performance, with area under the curve (AUC) of the receiver operating characteristics (ROC) of 0.84 in the independent hold-out cross-sectional dataset. High performance was also seen in external testing datasets for classifier 1 (AUC of 0.70 to 0.96). Classifier 2 that discriminates MCI from AD achieves AUC of 0.79 in the independent hold-out dataset and moderate to good performance in the external testing datasets (AUC of 0.56 to 0.93). The new data analysis methods, trained on 3T data, demonstrate potential for aiding AD early detection and disease progression on both 3T and 1.5T scanners.