Hybrid quaternion-based denoising and deep feature fusion for early Alzheimer's prognosis.
Authors
Affiliations (2)
Affiliations (2)
- Department of Computer Science and Engineering, Amrita School of Computing, Amrita Vishwa Vidyapeetham, Chennai, Tamil Nadu 601103 India.
- Department of Computer Science Engineering in Data Science, Dayananda Sagar Academy of Technology and Management, Udayapura, Kanakpura Road, Bangalore, 560082 India.
Abstract
Alzheimer's Disease (AD) is one of the most common neurodegenerative diseases that causes irreversible cognitive impairment and agnosia. Early detection is extremely valuable in diagnostics, but diagnostic procedures continue to rely on subjective approaches as well as traditional neuroimaging biomarkers, which frequently exhibit low tolerance to noise interference and imprecise structure localization. Traditional approaches also have difficulty combining multi-modal scans (MRI and PET) and detecting early-stage Alzheimer's disease characteristics (e.g., hippocampal atrophy). Some critical limitations of the current state-of-the-art include: (1) extreme sensitivity to noise; (2) poor spatio-temporal feature fusion; and (3) a lack of biomarker prioritization during disease progression. This problem is difficult to solve because the stages of Alzheimer's disease (AD) are not well differentiated. To address the aforementioned challenges, we propose a new study using the Hybrid Quaternion-Based Denoising and Deep Feature Fusion (HQD-FF) framework. The architecture will include a Quaternion Non-Local Means (QNLM) denoising procedure to suppress Rician/Rayleigh noise while keeping the structure intact. Following that, a hybrid CNN-RNN architecture will be used to extract spatial and temporal features using an attention-based fusion technique, allowing the identification of important biomarkers such as cortical thinning. The evaluation was carried out using the public multi-modal neuroimaging ADNI and OASIS datasets, as well as a private set of MRI and PET images. The experimental results show that HQD-FF achieves a PSNR of 23.47 dB and SSIM of 0.745, which is 12-15% better than traditional denoising techniques. HQD-FF also achieves 93.4% classification accuracy, outperforming current baselines by 6.2% in terms of AD diagnosis and stage differentiation (EMCI, LMCI, and AD). The proposed framework was also tested for robustness and statistical significance using ROC analysis, Area Under the Curve analysis, Confusion matrix analysis, and the Wilcoxon signed-rank statistical test. HQD-FF provides a robust, interpretable, and scalable framework for Alzheimer's diagnosis, with the potential to support early intervention and personalized medicine. The generalizable framework may pave the way for multimodal disease classification in broader neurodegenerative research.