International Conference on Recent Advances in Electrical, Electronics, Ubiquitous Communication, and Computational Intelligence · Apr 22, 2026
Hybrid MAMBA-Transformer Framework for Neurodegenerative Disease Screening from Brain Imaging
By Kamatchi S
Neurodegenerative diseases such as Alzheimer disease and Parkinson disease place a growing burden on patients and healthcare systems, yet early and reliable screening remains challenging. Structural brain MRI offers a non invasive window into subtle anatomical changes, but extracting clinically useful patterns from large scale cohorts is non trivial. Conventional 3D convolutional networks often struggle to capture long range dependencies across distant cortical and subcortical regions and can overfit to scanner specific artefacts. Pure transformer based models improve global context modelling but incur high computational cost and may be sensitive to limited data and noise. To address these limitations, this work proposes a Hybrid Mamba Transformer architecture that couples selective state space modelling with light weight self attention over volumetric MRI tokens. The model is designed to jointly encode fine grained local atrophy and global structural relationships while remaining efficient enough for practical screening workflows. The framework is evaluated on T1 weighted MRI scans drawn from four major cohorts, namely ADNI, AIBL, OASIS 3, and PPMI, covering cognitively normal controls, mild cognitive impairment, Alzheimer disease, and Parkinson disease. This diverse multi cohort setting allows us to study both in dataset performance and cross dataset generalisation under realistic domain shift. On the ADNI test set, the proposed model attains an accuracy of 94.3 percent with high sensitivity and specificity, and outperforms strong 3D CNN, transformer, and Mamba only baselines. Cross cohort experiments further show gains of about three to four percentage points in accuracy and improved AUC when transferring between ADNI, AIBL, and OASIS 3 compared with a Swin Transformer baseline. Overall, the results indicate that the Hybrid Mamba Transformer yields a compact yet expressive representation that is robust to data heterogeneity and tolerant to moderate image degradation. These properties suggest that the proposed approach is a promising candidate for integration into decision support pipelines for early neurodegenerative screening in routine clinical practice.