Pattern-Based Features Extraction Algorithm in Diagnosis of Neurodegenerative Diseases from Diffusion Magnetic Resonance Imaging

During the image classification process, feature extraction is a critical step to a successful model, especially in medical imaging. A comprehensive feature set described general proprieties of the image of interest but could further infer distinguishable characteristics from other groups. Radiomic analysis is one of the most popular approach to extract features from computed tomography scan of brain tumour. However, in neurodegenerative diseases, this approach might fail because the region-of-interest is not well defined.Changes in neurodegenerative diseases might be subtle and multi-focuses, which might involve extensive regions when in the advanced stage. For example, the affected regions in Parkinson's disease can start from brainstem in the early stage and spread to substantia nigra and basal ganglia. In Alzheimer's disease hippocampus could initially be affected and subsequently expand to the cortical regions in later stage. Therefore, in the early stage, structural changes such as cortical atrophy in a single region is rarely seen. However, because of increased sensitivity, this subtle alteration in brain function might be detectable in diffusion magnetic resonance imaging.The aim of the proposal is to develop a novel feature extraction and selection algorithm, based on the spatial pattern of the disease of interest. Specifically, this proposed algorithm of feature extraction is firstly going to identify the spatial pattern of the target disease, and construct a distinguishable network by accumulating subtle changes in each affected brain region. Compared to radiomic approach, the extracted features in our proposal could reflect the characteristics of the disease and provide improved clinical interpretation for the estimated models. The extracted features will be sent to an optimized feature selection algorithm in order to remove non-informative features. The optimization of feature selection machine will be combined with statistical filters and iterative greedy algorithm. The preliminary result achieved 92% accuracy in validation at the diagnostic performance of Alzheimer’s disease. The validation accuracy maintained over 80% when patients with Mild Cognitive Impairment was added.In the project, we will further optimize this pattern-based feature extraction algorithm, feature selection approach and classifiers and subsequently applied to additional neurodegenerative diseases. In this three-year project, 40 patients with Parkinson’s disease and Mild Cognitive Impairment will be recruited for classification model validation, respectively. Ultimately, we will combine our local database with international neurodegenerative diseases database to evaluate the feasibility and demonstrate the diagnostic performance.

Project ID:PB10907-3104
External Project ID:MOST109-2221-E182-009-MY3