Brain Cerebrospinal Fluid variation in Dementia using Radiomic Features and Optimized Machine Learning Techniques

Objective: Dementia causes degeneration of neurons in brain. This disorder prognosis is difficult to understand in a non-invasive manner. This study is an attempt to understand CSF texture variations across different dementia severity stages using optimized machine learning model. Since this biomarker has not been fully explored for different severity changes of brain diseases. Methods: In this study, ADNI database is considered to attain normal (HC), EMCI, MCI, LMCI and AD images. The skull stripped images are segmented using multilevel Tsallis based moth flame optimization. Further, GLCM, ZM, LBP, ST and PHOG features are extracted to study pathological texture variation in CSF. Multi-subspace randomization and collaboration (MSRC) based feature selection method is used to attain significant feature. Finally, SVM performs multiclass classification. Findings: Results indicate that the proposed method precisely segmented CSF region for all considered images with high correlation values of 0.9. The MSRC method identifies appropriate rank for the considered feature vectors based on the performance in multiple subspaces. ANOVA test identifies optimal features set with a p-value < 0.001 for classification. It is noticed that SVM classifier classify HC (99%) , EMCI (97%), MCI (96%), LMCI (98%) and AD (96.1%) images with better prediction. Novelty: The novelty of this work is to integrate optimized thresholding model, texture analysis, and statistical validation to establish a measurable link between CSF biomarkers and dementia severity stages. Hence, it could offer better clinical insights for clinicians to make appropriate decisions regarding the progression of dementia severity stages and its underlying changes. Keywords: Dementia, Cerebrospinal fluid (CSF), Moth flame optimization, Texture feature, Multi-subspace randomization and collaboration (MSRC)