Stage-Specific Hybrid ROI Subset Learning for MRI - Based Alzheimer's Diagnosis

Jiang J., Gan H. S., Wang Z., SELVER M. A., Mukhtar H., Setiadi D. R. I. M.

29th International Conference on Computer Supported Cooperative Work in Design, CSCWD 2026, Fuzhou, China, 13 - 15 May 2026, pp.2424-2429, (Full Text)

  • Publication Type: Conference Paper / Full Text
  • Doi Number: 10.1109/cscwd68734.2026.11582192
  • City: Fuzhou
  • Country: China
  • Page Numbers: pp.2424-2429
  • Keywords: Alzheimer's disease, deep learning, hybrid learning, magnetic resonance imaging, region of interest
  • Dokuz Eylül University Affiliated: Yes

Abstract

Accurate MRI-based differentiation of Alzheimer's disease (AD), mild cognitive impairment (MCI), and cognitively normal (CN) individuals remains challenging due to subtle stage-dependent atrophy patterns and variability across imaging sites. This work proposes a stage-specific hybrid ROI-subset learning framework that decomposes the diagnostic task into three binary problems and discovers compact, discriminative, and non-redundant anatomical region subsets. T1-weighted MRIs from the ADNI1 cohort were preprocessed using standardized N4 correction, skull stripping, nonlinear registration, and atlas parcellation (Harvard- Oxford and Desikan-Killiany). Fixed-size 3D patches extracted from each ROI were encoded using a contrastively pretrained 3D ResNet to obtain harmonized region-level embeddings. A multi-objective sequential selection strategy, namely optimizing discrimination, minimizing redundancy, and achieving subset compactness, identified stable task-specific ROI groups. Lightweight calibrated classifiers were trained under a fully nested cross-validation scheme to eliminate information leakage. The proposed method achieved AUCs of 0.98 (AD vs. CN), 0.84 (AD vs. MCI), and 0.91 (MCI vs. CN), outperforming whole-brain 3D CNNs, transformers, and ROI-based baselines while significantly reducing inference cost. Cross-atlas consistency and frozen-parameter validation on AIBL demonstrated robust anatomical reproducibility and strong transportability. These results highlight compact ROI-subset learning as an efficient and interpretable alternative to end-to-end volumetric deep learning for MRI-based Alzheimer's assessment.